diff --git a/README.md b/README.md
index 990b8bd..c1d110e 100644
--- a/README.md
+++ b/README.md
@@ -1,2 +1,129 @@
-# mlflow
+# MLflow Utils
 
+MLflow integration utilities for the DaviesTechLabs AI/ML platform.
+
+## Installation
+
+```bash
+pip install -r requirements.txt
+```
+
+Or from Gitea:
+```bash
+pip install git+https://git.daviestechlabs.io/daviestechlabs/mlflow.git
+```
+
+## Modules
+
+| Module | Description |
+|--------|-------------|
+| `client.py` | MLflow client configuration and helpers |
+| `tracker.py` | General MLflowTracker for experiments |
+| `inference_tracker.py` | Async inference metrics for NATS handlers |
+| `model_registry.py` | Model Registry with KServe metadata |
+| `kfp_components.py` | Kubeflow Pipeline MLflow components |
+| `experiment_comparison.py` | Compare experiments and runs |
+| `cli.py` | Command-line interface |
+
+## Quick Start
+
+```python
+from mlflow_utils import get_mlflow_client, MLflowTracker
+
+# Simple tracking
+with MLflowTracker(experiment_name="my-experiment") as tracker:
+    tracker.log_params({"learning_rate": 0.001})
+    tracker.log_metrics({"accuracy": 0.95})
+```
+
+## Inference Tracking
+
+For NATS handlers (chat-handler, voice-assistant):
+
+```python
+from mlflow_utils import InferenceMetricsTracker
+from mlflow_utils.inference_tracker import InferenceMetrics
+
+tracker = InferenceMetricsTracker(
+    experiment_name="voice-assistant-prod",
+    batch_size=100,  # Batch metrics before logging
+)
+
+# During request handling
+metrics = InferenceMetrics(
+    request_id="uuid",
+    total_latency=1.5,
+    llm_latency=0.8,
+    input_tokens=150,
+    output_tokens=200,
+)
+await tracker.log_inference(metrics)
+```
+
+## Model Registry
+
+Register models with KServe deployment metadata:
+
+```python
+from mlflow_utils.model_registry import register_model_for_kserve
+
+register_model_for_kserve(
+    model_name="my-qlora-adapter",
+    model_uri="runs:/abc123/model",
+    kserve_runtime="kserve-vllm",
+    gpu_type="amd-strixhalo",
+)
+```
+
+## Kubeflow Components
+
+Use in KFP pipelines:
+
+```python
+from mlflow_utils.kfp_components import (
+    log_experiment_component,
+    register_model_component,
+)
+```
+
+## CLI
+
+```bash
+# List experiments
+python -m mlflow_utils.cli list-experiments
+
+# Compare runs
+python -m mlflow_utils.cli compare-runs --experiment "qlora-training"
+
+# Export metrics
+python -m mlflow_utils.cli export --run-id abc123 --output metrics.json
+```
+
+## Configuration
+
+| Environment Variable | Default | Description |
+|---------------------|---------|-------------|
+| `MLFLOW_TRACKING_URI` | `http://mlflow.mlflow.svc.cluster.local:80` | MLflow server |
+| `MLFLOW_EXPERIMENT_NAME` | `default` | Default experiment |
+| `MLFLOW_ENABLE_ASYNC` | `true` | Async logging for handlers |
+
+## Module Structure
+
+```
+mlflow_utils/
+├── __init__.py              # Public API
+├── client.py                # Connection management
+├── tracker.py               # General experiment tracker
+├── inference_tracker.py     # Async inference metrics
+├── model_registry.py        # Model registration + KServe
+├── kfp_components.py        # Kubeflow components
+├── experiment_comparison.py # Run comparison tools
+└── cli.py                   # Command-line interface
+```
+
+## Related
+
+- [handler-base](https://git.daviestechlabs.io/daviestechlabs/handler-base) - Uses inference tracker
+- [kubeflow](https://git.daviestechlabs.io/daviestechlabs/kubeflow) - KFP components
+- [argo](https://git.daviestechlabs.io/daviestechlabs/argo) - Training workflows
+- [homelab-design](https://git.daviestechlabs.io/daviestechlabs/homelab-design) - Architecture docs
diff --git a/mlflow_utils/__init__.py b/mlflow_utils/__init__.py
new file mode 100644
index 0000000..0ec37ef
--- /dev/null
+++ b/mlflow_utils/__init__.py
@@ -0,0 +1,40 @@
+"""
+MLflow Integration Utilities for LLM Workflows
+
+This module provides MLflow integration for:
+- Kubeflow Pipelines experiment tracking
+- Model Registry with KServe deployment metadata
+- Inference metrics logging from NATS handlers
+- Experiment comparison and analysis
+
+Configuration:
+  Set MLFLOW_TRACKING_URI environment variable or use defaults:
+  - In-cluster: http://mlflow.mlflow.svc.cluster.local:80
+  - External: https://mlflow.lab.daviestechlabs.io
+
+Usage:
+  from mlflow_utils import get_mlflow_client, MLflowTracker
+  from mlflow_utils.kfp_components import log_experiment_component
+  from mlflow_utils.model_registry import register_model_for_kserve
+  from mlflow_utils.inference_tracker import InferenceMetricsTracker
+"""
+
+from .client import (
+    get_mlflow_client,
+    get_tracking_uri,
+    ensure_experiment,
+    MLflowConfig,
+)
+from .tracker import MLflowTracker
+from .inference_tracker import InferenceMetricsTracker
+
+__all__ = [
+    "get_mlflow_client",
+    "get_tracking_uri", 
+    "ensure_experiment",
+    "MLflowConfig",
+    "MLflowTracker",
+    "InferenceMetricsTracker",
+]
+
+__version__ = "1.0.0"
diff --git a/mlflow_utils/cli.py b/mlflow_utils/cli.py
new file mode 100644
index 0000000..c535a5a
--- /dev/null
+++ b/mlflow_utils/cli.py
@@ -0,0 +1,371 @@
+#!/usr/bin/env python3
+"""
+MLflow Experiment CLI
+
+Command-line interface for querying and comparing MLflow experiments.
+
+Usage:
+    # Compare recent runs in an experiment
+    python -m mlflow_utils.cli compare --experiment chat-inference --runs 5
+    
+    # Get best run by metric
+    python -m mlflow_utils.cli best --experiment evaluation --metric eval.accuracy
+    
+    # Generate performance report
+    python -m mlflow_utils.cli report --service chat-handler --hours 24
+    
+    # Check model promotion criteria
+    python -m mlflow_utils.cli promote --model whisper-finetuned \\
+        --experiment voice-evaluation \\
+        --criteria "eval.accuracy>=0.9,total_latency_p95<=2.0"
+    
+    # List experiments
+    python -m mlflow_utils.cli list-experiments
+    
+    # Query runs
+    python -m mlflow_utils.cli query --experiment chat-inference \\
+        --filter "metrics.total_latency_mean < 1.0" --limit 10
+"""
+
+import argparse
+import json
+import sys
+from typing import Optional
+
+from .client import get_mlflow_client, health_check
+from .experiment_comparison import (
+    ExperimentAnalyzer,
+    compare_experiments,
+    promotion_recommendation,
+    get_inference_performance_report,
+)
+from .model_registry import (
+    list_model_versions,
+    get_production_model,
+    generate_kserve_yaml,
+)
+
+
+def cmd_health(args):
+    """Check MLflow connectivity."""
+    result = health_check()
+    print(json.dumps(result, indent=2))
+    sys.exit(0 if result["connected"] else 1)
+
+
+def cmd_list_experiments(args):
+    """List all experiments."""
+    client = get_mlflow_client(tracking_uri=args.tracking_uri)
+    experiments = client.search_experiments()
+    
+    print(f"{'ID':<10} {'Name':<40} {'Artifact Location'}")
+    print("-" * 80)
+    for exp in experiments:
+        print(f"{exp.experiment_id:<10} {exp.name:<40} {exp.artifact_location}")
+
+
+def cmd_compare(args):
+    """Compare recent runs in an experiment."""
+    analyzer = ExperimentAnalyzer(
+        args.experiment,
+        tracking_uri=args.tracking_uri
+    )
+    
+    if args.run_ids:
+        run_ids = args.run_ids.split(",")
+        comparison = analyzer.compare_runs(run_ids=run_ids)
+    else:
+        comparison = analyzer.compare_runs(n_recent=args.runs)
+    
+    if args.json:
+        print(json.dumps(comparison.to_dict(), indent=2, default=str))
+    else:
+        print(comparison.summary_table())
+
+
+def cmd_best(args):
+    """Find the best run by a metric."""
+    analyzer = ExperimentAnalyzer(
+        args.experiment,
+        tracking_uri=args.tracking_uri
+    )
+    
+    best_run = analyzer.get_best_run(
+        metric=args.metric,
+        minimize=args.minimize,
+        filter_string=args.filter or "",
+    )
+    
+    if not best_run:
+        print(f"No runs found with metric '{args.metric}'")
+        sys.exit(1)
+    
+    result = {
+        "run_id": best_run.info.run_id,
+        "run_name": best_run.info.run_name,
+        "metric_value": best_run.data.metrics.get(args.metric),
+        "all_metrics": dict(best_run.data.metrics),
+        "params": dict(best_run.data.params),
+    }
+    
+    if args.json:
+        print(json.dumps(result, indent=2))
+    else:
+        print(f"Best Run: {best_run.info.run_name or best_run.info.run_id}")
+        print(f"  {args.metric}: {result['metric_value']}")
+        print(f"  Run ID: {best_run.info.run_id}")
+
+
+def cmd_summary(args):
+    """Get metrics summary for an experiment."""
+    analyzer = ExperimentAnalyzer(
+        args.experiment,
+        tracking_uri=args.tracking_uri
+    )
+    
+    summary = analyzer.get_metrics_summary(
+        hours=args.hours,
+        metrics=args.metrics.split(",") if args.metrics else None,
+    )
+    
+    if args.json:
+        print(json.dumps(summary, indent=2))
+    else:
+        print(f"Metrics Summary for '{args.experiment}' (last {args.hours}h)")
+        print("=" * 60)
+        for metric, stats in sorted(summary.items()):
+            print(f"\n{metric}:")
+            for stat, value in stats.items():
+                print(f"  {stat}: {value:.4f}")
+
+
+def cmd_report(args):
+    """Generate an inference performance report."""
+    report = get_inference_performance_report(
+        service_name=args.service,
+        hours=args.hours,
+        tracking_uri=args.tracking_uri,
+    )
+    
+    if args.json:
+        print(json.dumps(report, indent=2))
+    else:
+        print(f"Performance Report: {report['service']}")
+        print(f"Period: Last {report['period_hours']} hours")
+        print(f"Generated: {report['generated_at']}")
+        print()
+        
+        if report["latency"]:
+            print("Latency Metrics:")
+            for metric, stats in report["latency"].items():
+                if "mean" in stats:
+                    print(f"  {metric}: {stats['mean']:.4f}s (p50: {stats.get('median', 'N/A')})")
+        
+        if report["rag"]:
+            print("\nRAG Usage:")
+            for metric, stats in report["rag"].items():
+                print(f"  {metric}: {stats.get('mean', 'N/A')}")
+        
+        if report["errors"]:
+            print("\nError Rates:")
+            for metric, stats in report["errors"].items():
+                print(f"  {metric}: {stats:.2f}%")
+
+
+def cmd_promote(args):
+    """Check model promotion criteria."""
+    # Parse criteria
+    criteria = {}
+    for criterion in args.criteria.split(","):
+        # Parse "metric>=value" or "metric<=value" etc.
+        for op in [">=", "<=", ">", "<"]:
+            if op in criterion:
+                metric, value = criterion.split(op)
+                criteria[metric.strip()] = (op, float(value.strip()))
+                break
+    
+    rec = promotion_recommendation(
+        model_name=args.model,
+        experiment_name=args.experiment,
+        criteria=criteria,
+        tracking_uri=args.tracking_uri,
+    )
+    
+    if args.json:
+        print(json.dumps(rec.to_dict(), indent=2))
+    else:
+        status = "✓ RECOMMENDED" if rec.recommended else "✗ NOT RECOMMENDED"
+        print(f"Model: {args.model}")
+        print(f"Status: {status}")
+        print("\nCriteria Evaluation:")
+        for reason in rec.reasons:
+            print(f"  {reason}")
+
+
+def cmd_query(args):
+    """Query runs with a filter."""
+    analyzer = ExperimentAnalyzer(
+        args.experiment,
+        tracking_uri=args.tracking_uri
+    )
+    
+    runs = analyzer.search_runs(
+        filter_string=args.filter or "",
+        max_results=args.limit,
+    )
+    
+    if args.json:
+        result = [
+            {
+                "run_id": r.info.run_id,
+                "run_name": r.info.run_name,
+                "status": r.info.status,
+                "metrics": dict(r.data.metrics),
+                "params": dict(r.data.params),
+            }
+            for r in runs
+        ]
+        print(json.dumps(result, indent=2))
+    else:
+        print(f"Found {len(runs)} runs")
+        for run in runs:
+            print(f"\n{run.info.run_name or run.info.run_id}")
+            print(f"  ID: {run.info.run_id}")
+            print(f"  Status: {run.info.status}")
+
+
+def cmd_models(args):
+    """List registered models."""
+    client = get_mlflow_client(tracking_uri=args.tracking_uri)
+    
+    if args.model:
+        versions = list_model_versions(args.model, tracking_uri=args.tracking_uri)
+        
+        if args.json:
+            print(json.dumps(versions, indent=2, default=str))
+        else:
+            print(f"Model: {args.model}")
+            for v in versions:
+                print(f"  v{v['version']} ({v['stage']}): {v['description'][:50] if v['description'] else 'No description'}")
+    else:
+        # List all models
+        models = client.search_registered_models()
+        
+        if args.json:
+            result = [{"name": m.name, "description": m.description} for m in models]
+            print(json.dumps(result, indent=2))
+        else:
+            print(f"{'Model Name':<40} Description")
+            print("-" * 80)
+            for model in models:
+                desc = (model.description or "")[:35]
+                print(f"{model.name:<40} {desc}")
+
+
+def cmd_kserve(args):
+    """Generate KServe manifest for a model."""
+    yaml_str = generate_kserve_yaml(
+        model_name=args.model,
+        version=args.version,
+        namespace=args.namespace,
+        output_path=args.output,
+        tracking_uri=args.tracking_uri,
+    )
+    
+    if not args.output:
+        print(yaml_str)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="MLflow Experiment CLI",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    
+    parser.add_argument(
+        "--tracking-uri",
+        default=None,
+        help="MLflow tracking URI (default: from env or in-cluster)",
+    )
+    parser.add_argument(
+        "--json",
+        action="store_true",
+        help="Output as JSON",
+    )
+    
+    subparsers = parser.add_subparsers(dest="command", help="Commands")
+    
+    # health
+    health_parser = subparsers.add_parser("health", help="Check MLflow connectivity")
+    health_parser.set_defaults(func=cmd_health)
+    
+    # list-experiments
+    list_parser = subparsers.add_parser("list-experiments", help="List experiments")
+    list_parser.set_defaults(func=cmd_list_experiments)
+    
+    # compare
+    compare_parser = subparsers.add_parser("compare", help="Compare runs")
+    compare_parser.add_argument("--experiment", "-e", required=True, help="Experiment name")
+    compare_parser.add_argument("--runs", "-n", type=int, default=5, help="Number of recent runs")
+    compare_parser.add_argument("--run-ids", help="Comma-separated run IDs to compare")
+    compare_parser.set_defaults(func=cmd_compare)
+    
+    # best
+    best_parser = subparsers.add_parser("best", help="Find best run by metric")
+    best_parser.add_argument("--experiment", "-e", required=True, help="Experiment name")
+    best_parser.add_argument("--metric", "-m", required=True, help="Metric to optimize")
+    best_parser.add_argument("--minimize", action="store_true", help="Minimize metric (default: maximize)")
+    best_parser.add_argument("--filter", "-f", help="Filter string")
+    best_parser.set_defaults(func=cmd_best)
+    
+    # summary
+    summary_parser = subparsers.add_parser("summary", help="Get metrics summary")
+    summary_parser.add_argument("--experiment", "-e", required=True, help="Experiment name")
+    summary_parser.add_argument("--hours", type=int, default=24, help="Hours of data")
+    summary_parser.add_argument("--metrics", help="Comma-separated metric names")
+    summary_parser.set_defaults(func=cmd_summary)
+    
+    # report
+    report_parser = subparsers.add_parser("report", help="Generate performance report")
+    report_parser.add_argument("--service", "-s", required=True, help="Service name")
+    report_parser.add_argument("--hours", type=int, default=24, help="Hours of data")
+    report_parser.set_defaults(func=cmd_report)
+    
+    # promote
+    promote_parser = subparsers.add_parser("promote", help="Check promotion criteria")
+    promote_parser.add_argument("--model", "-m", required=True, help="Model name")
+    promote_parser.add_argument("--experiment", "-e", required=True, help="Experiment name")
+    promote_parser.add_argument("--criteria", "-c", required=True, help="Criteria (e.g., 'accuracy>=0.9,latency<=2.0')")
+    promote_parser.set_defaults(func=cmd_promote)
+    
+    # query
+    query_parser = subparsers.add_parser("query", help="Query runs")
+    query_parser.add_argument("--experiment", "-e", required=True, help="Experiment name")
+    query_parser.add_argument("--filter", "-f", help="MLflow filter string")
+    query_parser.add_argument("--limit", "-l", type=int, default=20, help="Max results")
+    query_parser.set_defaults(func=cmd_query)
+    
+    # models
+    models_parser = subparsers.add_parser("models", help="List registered models")
+    models_parser.add_argument("--model", "-m", help="Specific model name")
+    models_parser.set_defaults(func=cmd_models)
+    
+    # kserve
+    kserve_parser = subparsers.add_parser("kserve", help="Generate KServe manifest")
+    kserve_parser.add_argument("--model", "-m", required=True, help="Model name")
+    kserve_parser.add_argument("--version", "-v", type=int, help="Model version")
+    kserve_parser.add_argument("--namespace", "-n", default="ai-ml", help="K8s namespace")
+    kserve_parser.add_argument("--output", "-o", help="Output file path")
+    kserve_parser.set_defaults(func=cmd_kserve)
+    
+    args = parser.parse_args()
+    
+    if not args.command:
+        parser.print_help()
+        sys.exit(1)
+    
+    args.func(args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/mlflow_utils/client.py b/mlflow_utils/client.py
new file mode 100644
index 0000000..b448d99
--- /dev/null
+++ b/mlflow_utils/client.py
@@ -0,0 +1,209 @@
+"""
+MLflow Client Configuration and Initialization
+
+Provides a configured MLflow client for all integrations in the LLM workflows.
+Supports both in-cluster and external access patterns.
+"""
+
+import os
+import logging
+from dataclasses import dataclass, field
+from typing import Optional, Dict, Any
+
+import mlflow
+from mlflow.tracking import MlflowClient
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class MLflowConfig:
+    """Configuration for MLflow integration."""
+    
+    # Tracking server URIs
+    tracking_uri: str = field(
+        default_factory=lambda: os.environ.get(
+            "MLFLOW_TRACKING_URI",
+            "http://mlflow.mlflow.svc.cluster.local:80"
+        )
+    )
+    external_uri: str = field(
+        default_factory=lambda: os.environ.get(
+            "MLFLOW_EXTERNAL_URI",
+            "https://mlflow.lab.daviestechlabs.io"
+        )
+    )
+    
+    # Artifact storage (NFS PVC mount)
+    artifact_location: str = field(
+        default_factory=lambda: os.environ.get(
+            "MLFLOW_ARTIFACT_LOCATION",
+            "/mlflow/artifacts"
+        )
+    )
+    
+    # Default experiment settings
+    default_experiment: str = field(
+        default_factory=lambda: os.environ.get(
+            "MLFLOW_DEFAULT_EXPERIMENT",
+            "llm-workflows"
+        )
+    )
+    
+    # Service identification
+    service_name: str = field(
+        default_factory=lambda: os.environ.get(
+            "OTEL_SERVICE_NAME",
+            "unknown-service"
+        )
+    )
+    
+    # Additional tags to add to all runs
+    default_tags: Dict[str, str] = field(default_factory=dict)
+    
+    def __post_init__(self):
+        """Add default tags based on environment."""
+        env_tags = {
+            "environment": os.environ.get("DEPLOYMENT_ENV", "production"),
+            "hostname": os.environ.get("HOSTNAME", "unknown"),
+            "namespace": os.environ.get("OTEL_SERVICE_NAMESPACE", "ai-ml"),
+        }
+        self.default_tags = {**env_tags, **self.default_tags}
+
+
+def get_tracking_uri(external: bool = False) -> str:
+    """
+    Get the appropriate MLflow tracking URI.
+    
+    Args:
+        external: If True, return the external URI for outside-cluster access
+        
+    Returns:
+        The MLflow tracking URI string
+    """
+    config = MLflowConfig()
+    return config.external_uri if external else config.tracking_uri
+
+
+def get_mlflow_client(
+    tracking_uri: Optional[str] = None,
+    configure_global: bool = True
+) -> MlflowClient:
+    """
+    Get a configured MLflow client.
+    
+    Args:
+        tracking_uri: Override the default tracking URI
+        configure_global: If True, also set mlflow.set_tracking_uri()
+        
+    Returns:
+        Configured MlflowClient instance
+    """
+    uri = tracking_uri or get_tracking_uri()
+    
+    if configure_global:
+        mlflow.set_tracking_uri(uri)
+        logger.info(f"MLflow tracking URI set to: {uri}")
+    
+    client = MlflowClient(tracking_uri=uri)
+    return client
+
+
+def ensure_experiment(
+    experiment_name: str,
+    artifact_location: Optional[str] = None,
+    tags: Optional[Dict[str, str]] = None
+) -> str:
+    """
+    Ensure an experiment exists, creating it if necessary.
+    
+    Args:
+        experiment_name: Name of the experiment
+        artifact_location: Override default artifact location
+        tags: Additional tags for the experiment
+        
+    Returns:
+        The experiment ID
+    """
+    config = MLflowConfig()
+    client = get_mlflow_client()
+    
+    # Check if experiment exists
+    experiment = client.get_experiment_by_name(experiment_name)
+    
+    if experiment is None:
+        # Create the experiment
+        artifact_loc = artifact_location or f"{config.artifact_location}/{experiment_name}"
+        experiment_id = client.create_experiment(
+            name=experiment_name,
+            artifact_location=artifact_loc,
+            tags=tags or {}
+        )
+        logger.info(f"Created experiment '{experiment_name}' with ID: {experiment_id}")
+    else:
+        experiment_id = experiment.experiment_id
+        logger.debug(f"Using existing experiment '{experiment_name}' with ID: {experiment_id}")
+    
+    return experiment_id
+
+
+def get_or_create_registered_model(
+    model_name: str,
+    description: Optional[str] = None,
+    tags: Optional[Dict[str, str]] = None
+) -> str:
+    """
+    Get or create a registered model in the Model Registry.
+    
+    Args:
+        model_name: Name of the model to register
+        description: Model description
+        tags: Tags for the model
+        
+    Returns:
+        The registered model name
+    """
+    client = get_mlflow_client()
+    
+    try:
+        # Check if model exists
+        client.get_registered_model(model_name)
+        logger.debug(f"Using existing registered model: {model_name}")
+    except mlflow.exceptions.MlflowException:
+        # Create the model
+        client.create_registered_model(
+            name=model_name,
+            description=description or f"Model for {model_name}",
+            tags=tags or {}
+        )
+        logger.info(f"Created registered model: {model_name}")
+    
+    return model_name
+
+
+def health_check() -> Dict[str, Any]:
+    """
+    Check MLflow server connectivity and return status.
+    
+    Returns:
+        Dictionary with health status information
+    """
+    config = MLflowConfig()
+    result = {
+        "tracking_uri": config.tracking_uri,
+        "external_uri": config.external_uri,
+        "connected": False,
+        "error": None,
+    }
+    
+    try:
+        client = get_mlflow_client(configure_global=False)
+        # Try to list experiments as a health check
+        experiments = client.search_experiments(max_results=1)
+        result["connected"] = True
+        result["experiment_count"] = len(experiments)
+    except Exception as e:
+        result["error"] = str(e)
+        logger.error(f"MLflow health check failed: {e}")
+    
+    return result
diff --git a/mlflow_utils/experiment_comparison.py b/mlflow_utils/experiment_comparison.py
new file mode 100644
index 0000000..f66d164
--- /dev/null
+++ b/mlflow_utils/experiment_comparison.py
@@ -0,0 +1,664 @@
+"""
+Experiment Comparison and Analysis Utilities
+
+Provides tools for comparing model versions, querying experiments,
+and making data-driven decisions about model promotion to production.
+
+Features:
+- Compare multiple runs/experiments side by side
+- Query experiments by tags, metrics, or parameters
+- Analyze inference metrics from NATS handlers
+- Generate promotion recommendations
+- Export comparison reports
+
+Usage:
+    from mlflow_utils.experiment_comparison import (
+        ExperimentAnalyzer,
+        compare_runs,
+        get_best_run,
+        promotion_recommendation,
+    )
+    
+    analyzer = ExperimentAnalyzer("chat-inference")
+    
+    # Compare last N runs
+    comparison = analyzer.compare_recent_runs(n=5)
+    
+    # Find best performing model
+    best = analyzer.get_best_run(metric="total_latency_mean", minimize=True)
+    
+    # Get promotion recommendation
+    rec = analyzer.promotion_recommendation(
+        model_name="whisper-finetuned",
+        min_accuracy=0.9,
+        max_latency_p95=2.0
+    )
+"""
+
+import os
+import json
+import logging
+from datetime import datetime, timedelta
+from typing import Optional, Dict, Any, List, Tuple, Union
+from dataclasses import dataclass, field
+from collections import defaultdict
+
+import mlflow
+from mlflow.tracking import MlflowClient
+from mlflow.entities import Run, Experiment
+
+from .client import get_mlflow_client, MLflowConfig
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class RunComparison:
+    """Comparison result for multiple MLflow runs."""
+    run_ids: List[str]
+    experiment_name: str
+    
+    # Metric comparisons (metric_name -> {run_id -> value})
+    metrics: Dict[str, Dict[str, float]] = field(default_factory=dict)
+    
+    # Parameter differences
+    params: Dict[str, Dict[str, str]] = field(default_factory=dict)
+    
+    # Run metadata
+    run_names: Dict[str, str] = field(default_factory=dict)
+    start_times: Dict[str, datetime] = field(default_factory=dict)
+    durations: Dict[str, float] = field(default_factory=dict)
+    
+    # Best performers by metric
+    best_by_metric: Dict[str, str] = field(default_factory=dict)
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for serialization."""
+        return {
+            "run_ids": self.run_ids,
+            "experiment_name": self.experiment_name,
+            "metrics": self.metrics,
+            "params": self.params,
+            "run_names": self.run_names,
+            "best_by_metric": self.best_by_metric,
+        }
+    
+    def summary_table(self) -> str:
+        """Generate a text summary table of the comparison."""
+        if not self.run_ids:
+            return "No runs to compare"
+        
+        lines = []
+        lines.append(f"Experiment: {self.experiment_name}")
+        lines.append(f"Comparing {len(self.run_ids)} runs")
+        lines.append("")
+        
+        # Header
+        header = ["Metric"] + [self.run_names.get(rid, rid[:8]) for rid in self.run_ids]
+        lines.append(" | ".join(header))
+        lines.append("-" * (len(lines[-1]) + 10))
+        
+        # Metrics
+        for metric_name, values in sorted(self.metrics.items()):
+            row = [metric_name]
+            for run_id in self.run_ids:
+                value = values.get(run_id)
+                if value is not None:
+                    row.append(f"{value:.4f}")
+                else:
+                    row.append("N/A")
+            lines.append(" | ".join(row))
+        
+        return "\n".join(lines)
+
+
+@dataclass
+class PromotionRecommendation:
+    """Recommendation for model promotion."""
+    model_name: str
+    version: Optional[int]
+    recommended: bool
+    reasons: List[str]
+    metrics_summary: Dict[str, float]
+    comparison_with_production: Optional[Dict[str, Any]] = None
+    
+    def to_dict(self) -> Dict[str, Any]:
+        return {
+            "model_name": self.model_name,
+            "version": self.version,
+            "recommended": self.recommended,
+            "reasons": self.reasons,
+            "metrics_summary": self.metrics_summary,
+            "comparison_with_production": self.comparison_with_production,
+        }
+
+
+class ExperimentAnalyzer:
+    """
+    Analyze MLflow experiments for model comparison and promotion decisions.
+    
+    Example:
+        analyzer = ExperimentAnalyzer("chat-inference")
+        
+        # Get metrics summary for last 24 hours
+        summary = analyzer.get_metrics_summary(hours=24)
+        
+        # Compare models by accuracy
+        best = analyzer.get_best_run(metric="eval.accuracy", minimize=False)
+        
+        # Analyze inference latency trends
+        trends = analyzer.get_metric_trends("total_latency_mean", days=7)
+    """
+    
+    def __init__(
+        self,
+        experiment_name: str,
+        tracking_uri: Optional[str] = None,
+    ):
+        """
+        Initialize the experiment analyzer.
+        
+        Args:
+            experiment_name: Name of the MLflow experiment to analyze
+            tracking_uri: Override default tracking URI
+        """
+        self.experiment_name = experiment_name
+        self.tracking_uri = tracking_uri
+        self.client = get_mlflow_client(tracking_uri=tracking_uri)
+        self._experiment: Optional[Experiment] = None
+    
+    @property
+    def experiment(self) -> Optional[Experiment]:
+        """Get the experiment object, fetching if needed."""
+        if self._experiment is None:
+            self._experiment = self.client.get_experiment_by_name(self.experiment_name)
+        return self._experiment
+    
+    def search_runs(
+        self,
+        filter_string: str = "",
+        order_by: Optional[List[str]] = None,
+        max_results: int = 100,
+        run_view_type: str = "ACTIVE_ONLY",
+    ) -> List[Run]:
+        """
+        Search for runs matching criteria.
+        
+        Args:
+            filter_string: MLflow search filter (e.g., "metrics.accuracy > 0.9")
+            order_by: List of order clauses (e.g., ["metrics.accuracy DESC"])
+            max_results: Maximum runs to return
+            run_view_type: ACTIVE_ONLY, DELETED_ONLY, or ALL
+            
+        Returns:
+            List of matching Run objects
+        """
+        if not self.experiment:
+            logger.warning(f"Experiment '{self.experiment_name}' not found")
+            return []
+        
+        runs = self.client.search_runs(
+            experiment_ids=[self.experiment.experiment_id],
+            filter_string=filter_string,
+            order_by=order_by or ["start_time DESC"],
+            max_results=max_results,
+        )
+        
+        return runs
+    
+    def get_recent_runs(
+        self,
+        n: int = 10,
+        hours: Optional[int] = None,
+    ) -> List[Run]:
+        """
+        Get the most recent runs.
+        
+        Args:
+            n: Number of runs to return
+            hours: Only include runs from the last N hours
+            
+        Returns:
+            List of Run objects
+        """
+        filter_string = ""
+        if hours:
+            cutoff = datetime.now() - timedelta(hours=hours)
+            cutoff_ms = int(cutoff.timestamp() * 1000)
+            filter_string = f"attributes.start_time >= {cutoff_ms}"
+        
+        return self.search_runs(
+            filter_string=filter_string,
+            order_by=["start_time DESC"],
+            max_results=n,
+        )
+    
+    def compare_runs(
+        self,
+        run_ids: Optional[List[str]] = None,
+        n_recent: int = 5,
+    ) -> RunComparison:
+        """
+        Compare multiple runs side by side.
+        
+        Args:
+            run_ids: Specific run IDs to compare, or None for recent runs
+            n_recent: If run_ids is None, compare this many recent runs
+            
+        Returns:
+            RunComparison object with detailed comparison
+        """
+        if run_ids:
+            runs = [self.client.get_run(rid) for rid in run_ids]
+        else:
+            runs = self.get_recent_runs(n=n_recent)
+        
+        comparison = RunComparison(
+            run_ids=[r.info.run_id for r in runs],
+            experiment_name=self.experiment_name,
+        )
+        
+        # Collect all metrics and find best performers
+        all_metrics: Dict[str, Dict[str, float]] = defaultdict(dict)
+        
+        for run in runs:
+            run_id = run.info.run_id
+            
+            # Metadata
+            comparison.run_names[run_id] = run.info.run_name or run_id[:8]
+            comparison.start_times[run_id] = datetime.fromtimestamp(
+                run.info.start_time / 1000
+            )
+            if run.info.end_time:
+                comparison.durations[run_id] = (
+                    run.info.end_time - run.info.start_time
+                ) / 1000
+            
+            # Metrics
+            for key, value in run.data.metrics.items():
+                all_metrics[key][run_id] = value
+            
+            # Params
+            for key, value in run.data.params.items():
+                if key not in comparison.params:
+                    comparison.params[key] = {}
+                comparison.params[key][run_id] = value
+        
+        comparison.metrics = dict(all_metrics)
+        
+        # Find best performers for each metric
+        for metric_name, values in all_metrics.items():
+            if not values:
+                continue
+            
+            # Determine if lower is better based on metric name
+            minimize = any(
+                term in metric_name.lower()
+                for term in ["latency", "error", "loss", "time"]
+            )
+            
+            if minimize:
+                best_id = min(values.keys(), key=lambda k: values[k])
+            else:
+                best_id = max(values.keys(), key=lambda k: values[k])
+            
+            comparison.best_by_metric[metric_name] = best_id
+        
+        return comparison
+    
+    def get_best_run(
+        self,
+        metric: str,
+        minimize: bool = True,
+        filter_string: str = "",
+        max_results: int = 100,
+    ) -> Optional[Run]:
+        """
+        Get the best run by a specific metric.
+        
+        Args:
+            metric: Metric name to optimize
+            minimize: If True, find minimum; if False, find maximum
+            filter_string: Additional filter criteria
+            max_results: Maximum runs to consider
+            
+        Returns:
+            Best Run object, or None if no runs found
+        """
+        direction = "ASC" if minimize else "DESC"
+        
+        runs = self.search_runs(
+            filter_string=filter_string,
+            order_by=[f"metrics.{metric} {direction}"],
+            max_results=max_results,
+        )
+        
+        # Filter to only runs that have the metric
+        runs_with_metric = [
+            r for r in runs
+            if metric in r.data.metrics
+        ]
+        
+        return runs_with_metric[0] if runs_with_metric else None
+    
+    def get_metrics_summary(
+        self,
+        hours: Optional[int] = None,
+        metrics: Optional[List[str]] = None,
+    ) -> Dict[str, Dict[str, float]]:
+        """
+        Get summary statistics for metrics.
+        
+        Args:
+            hours: Only include runs from the last N hours
+            metrics: Specific metrics to summarize (None for all)
+            
+        Returns:
+            Dict mapping metric names to {mean, min, max, count}
+        """
+        import statistics
+        
+        runs = self.get_recent_runs(n=1000, hours=hours)
+        
+        # Collect all metric values
+        metric_values: Dict[str, List[float]] = defaultdict(list)
+        
+        for run in runs:
+            for key, value in run.data.metrics.items():
+                if metrics is None or key in metrics:
+                    metric_values[key].append(value)
+        
+        # Calculate statistics
+        summary = {}
+        for metric_name, values in metric_values.items():
+            if not values:
+                continue
+            
+            summary[metric_name] = {
+                "mean": statistics.mean(values),
+                "min": min(values),
+                "max": max(values),
+                "count": len(values),
+            }
+            
+            if len(values) >= 2:
+                summary[metric_name]["stdev"] = statistics.stdev(values)
+                summary[metric_name]["median"] = statistics.median(values)
+        
+        return summary
+    
+    def get_metric_trends(
+        self,
+        metric: str,
+        days: int = 7,
+        granularity_hours: int = 1,
+    ) -> List[Dict[str, Any]]:
+        """
+        Get metric trends over time.
+        
+        Args:
+            metric: Metric name to track
+            days: Number of days to look back
+            granularity_hours: Time bucket size in hours
+            
+        Returns:
+            List of {timestamp, mean, min, max, count} dicts
+        """
+        import statistics
+        
+        runs = self.get_recent_runs(n=10000, hours=days * 24)
+        
+        # Group runs by time bucket
+        buckets: Dict[int, List[float]] = defaultdict(list)
+        bucket_size_ms = granularity_hours * 3600 * 1000
+        
+        for run in runs:
+            if metric not in run.data.metrics:
+                continue
+            
+            bucket = (run.info.start_time // bucket_size_ms) * bucket_size_ms
+            buckets[bucket].append(run.data.metrics[metric])
+        
+        # Calculate statistics per bucket
+        trends = []
+        for bucket_ts, values in sorted(buckets.items()):
+            trend = {
+                "timestamp": datetime.fromtimestamp(bucket_ts / 1000).isoformat(),
+                "count": len(values),
+                "mean": statistics.mean(values),
+                "min": min(values),
+                "max": max(values),
+            }
+            if len(values) >= 2:
+                trend["stdev"] = statistics.stdev(values)
+            trends.append(trend)
+        
+        return trends
+    
+    def get_runs_by_tag(
+        self,
+        tag_key: str,
+        tag_value: str,
+        max_results: int = 100,
+    ) -> List[Run]:
+        """
+        Get runs with a specific tag.
+        
+        Args:
+            tag_key: Tag key to filter by
+            tag_value: Tag value to match
+            max_results: Maximum runs to return
+            
+        Returns:
+            List of matching Run objects
+        """
+        return self.search_runs(
+            filter_string=f"tags.{tag_key} = '{tag_value}'",
+            max_results=max_results,
+        )
+    
+    def get_model_runs(
+        self,
+        model_name: str,
+        max_results: int = 100,
+    ) -> List[Run]:
+        """
+        Get runs for a specific model.
+        
+        Args:
+            model_name: Model name to filter by
+            max_results: Maximum runs to return
+            
+        Returns:
+            List of matching Run objects
+        """
+        # Try different tag conventions
+        runs = self.search_runs(
+            filter_string=f"tags.`model.name` = '{model_name}'",
+            max_results=max_results,
+        )
+        
+        if not runs:
+            # Try params
+            runs = self.search_runs(
+                filter_string=f"params.model_name = '{model_name}'",
+                max_results=max_results,
+            )
+        
+        return runs
+
+
+def compare_experiments(
+    experiment_names: List[str],
+    metric: str,
+    tracking_uri: Optional[str] = None,
+) -> Dict[str, Dict[str, float]]:
+    """
+    Compare metrics across multiple experiments.
+    
+    Args:
+        experiment_names: Names of experiments to compare
+        metric: Metric to compare
+        tracking_uri: Override default tracking URI
+        
+    Returns:
+        Dict mapping experiment names to metric statistics
+    """
+    results = {}
+    
+    for exp_name in experiment_names:
+        analyzer = ExperimentAnalyzer(exp_name, tracking_uri=tracking_uri)
+        summary = analyzer.get_metrics_summary(metrics=[metric])
+        if metric in summary:
+            results[exp_name] = summary[metric]
+    
+    return results
+
+
+def promotion_recommendation(
+    model_name: str,
+    experiment_name: str,
+    criteria: Dict[str, Tuple[str, float]],
+    tracking_uri: Optional[str] = None,
+) -> PromotionRecommendation:
+    """
+    Generate a recommendation for model promotion.
+    
+    Args:
+        model_name: Name of the model to evaluate
+        experiment_name: Experiment containing evaluation runs
+        criteria: Dict of {metric: (comparison, threshold)}
+                  comparison is one of: ">=", "<=", ">", "<"
+                  e.g., {"eval.accuracy": (">=", 0.9), "total_latency_p95": ("<=", 2.0)}
+        tracking_uri: Override default tracking URI
+        
+    Returns:
+        PromotionRecommendation with decision and reasons
+    """
+    analyzer = ExperimentAnalyzer(experiment_name, tracking_uri=tracking_uri)
+    
+    # Get model runs
+    runs = analyzer.get_model_runs(model_name, max_results=10)
+    
+    if not runs:
+        return PromotionRecommendation(
+            model_name=model_name,
+            version=None,
+            recommended=False,
+            reasons=["No runs found for this model"],
+            metrics_summary={},
+        )
+    
+    # Get the most recent run
+    latest_run = runs[0]
+    metrics = latest_run.data.metrics
+    
+    # Evaluate criteria
+    reasons = []
+    passed = True
+    
+    comparisons = {
+        ">=": lambda a, b: a >= b,
+        "<=": lambda a, b: a <= b,
+        ">": lambda a, b: a > b,
+        "<": lambda a, b: a < b,
+    }
+    
+    for metric_name, (comparison, threshold) in criteria.items():
+        if metric_name not in metrics:
+            reasons.append(f"Metric '{metric_name}' not found")
+            passed = False
+            continue
+        
+        value = metrics[metric_name]
+        compare_fn = comparisons.get(comparison)
+        
+        if compare_fn is None:
+            reasons.append(f"Invalid comparison operator: {comparison}")
+            continue
+        
+        if compare_fn(value, threshold):
+            reasons.append(f"✓ {metric_name}: {value:.4f} {comparison} {threshold}")
+        else:
+            reasons.append(f"✗ {metric_name}: {value:.4f} NOT {comparison} {threshold}")
+            passed = False
+    
+    # Extract version from tags if available
+    version = None
+    if "mlflow.version" in latest_run.data.tags:
+        try:
+            version = int(latest_run.data.tags["mlflow.version"])
+        except ValueError:
+            pass
+    
+    return PromotionRecommendation(
+        model_name=model_name,
+        version=version,
+        recommended=passed,
+        reasons=reasons,
+        metrics_summary=dict(metrics),
+    )
+
+
+def get_inference_performance_report(
+    service_name: str = "chat-handler",
+    hours: int = 24,
+    tracking_uri: Optional[str] = None,
+) -> Dict[str, Any]:
+    """
+    Generate an inference performance report for a service.
+    
+    Args:
+        service_name: Service name (chat-handler, voice-assistant)
+        hours: Hours of data to analyze
+        tracking_uri: Override default tracking URI
+        
+    Returns:
+        Performance report dictionary
+    """
+    experiment_name = f"{service_name.replace('-', '')}-inference"
+    analyzer = ExperimentAnalyzer(experiment_name, tracking_uri=tracking_uri)
+    
+    # Get summary metrics
+    summary = analyzer.get_metrics_summary(hours=hours)
+    
+    # Key latency metrics
+    latency_metrics = [
+        "total_latency_mean",
+        "total_latency_p50",
+        "total_latency_p95",
+        "llm_latency_mean",
+        "embedding_latency_mean",
+        "rag_search_latency_mean",
+    ]
+    
+    report = {
+        "service": service_name,
+        "period_hours": hours,
+        "generated_at": datetime.now().isoformat(),
+        "latency": {},
+        "throughput": {},
+        "rag": {},
+        "errors": {},
+    }
+    
+    # Latency section
+    for metric in latency_metrics:
+        if metric in summary:
+            report["latency"][metric] = summary[metric]
+    
+    # Throughput
+    if "total_requests" in summary:
+        report["throughput"]["total_requests"] = summary["total_requests"]["mean"]
+    
+    # RAG usage
+    rag_metrics = ["rag_enabled_pct", "rag_documents_retrieved_mean", "rag_documents_used_mean"]
+    for metric in rag_metrics:
+        if metric in summary:
+            report["rag"][metric] = summary[metric]
+    
+    # Error rate
+    if "error_rate" in summary:
+        report["errors"]["error_rate_pct"] = summary["error_rate"]["mean"]
+    
+    return report
diff --git a/mlflow_utils/inference_tracker.py b/mlflow_utils/inference_tracker.py
new file mode 100644
index 0000000..820db8a
--- /dev/null
+++ b/mlflow_utils/inference_tracker.py
@@ -0,0 +1,431 @@
+"""
+Inference Metrics Tracker for NATS Handlers
+
+Provides async-compatible MLflow logging for real-time inference
+metrics from chat-handler and voice-assistant services.
+
+Designed to integrate with the existing OpenTelemetry setup and
+complement OTel metrics with MLflow experiment tracking for
+longer-term analysis and model comparison.
+"""
+
+import os
+import time
+import asyncio
+import logging
+from typing import Optional, Dict, Any, List
+from dataclasses import dataclass, field
+from collections import defaultdict
+from concurrent.futures import ThreadPoolExecutor
+from functools import partial
+
+import mlflow
+from mlflow.tracking import MlflowClient
+
+from .client import get_mlflow_client, ensure_experiment, MLflowConfig
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class InferenceMetrics:
+    """Metrics collected during an inference request."""
+    request_id: str
+    user_id: Optional[str] = None
+    session_id: Optional[str] = None
+    
+    # Timing metrics (in seconds)
+    total_latency: float = 0.0
+    embedding_latency: float = 0.0
+    rag_search_latency: float = 0.0
+    rerank_latency: float = 0.0
+    llm_latency: float = 0.0
+    tts_latency: float = 0.0
+    stt_latency: float = 0.0
+    
+    # Token/size metrics
+    input_tokens: int = 0
+    output_tokens: int = 0
+    total_tokens: int = 0
+    prompt_length: int = 0
+    response_length: int = 0
+    
+    # RAG metrics
+    rag_enabled: bool = False
+    rag_documents_retrieved: int = 0
+    rag_documents_used: int = 0
+    reranker_enabled: bool = False
+    
+    # Quality indicators
+    is_streaming: bool = False
+    is_premium: bool = False
+    has_error: bool = False
+    error_message: Optional[str] = None
+    
+    # Model information
+    model_name: Optional[str] = None
+    model_endpoint: Optional[str] = None
+    
+    # Timestamps
+    timestamp: float = field(default_factory=time.time)
+    
+    def as_metrics_dict(self) -> Dict[str, float]:
+        """Convert numeric fields to a metrics dictionary."""
+        return {
+            "total_latency": self.total_latency,
+            "embedding_latency": self.embedding_latency,
+            "rag_search_latency": self.rag_search_latency,
+            "rerank_latency": self.rerank_latency,
+            "llm_latency": self.llm_latency,
+            "tts_latency": self.tts_latency,
+            "stt_latency": self.stt_latency,
+            "input_tokens": float(self.input_tokens),
+            "output_tokens": float(self.output_tokens),
+            "total_tokens": float(self.total_tokens),
+            "prompt_length": float(self.prompt_length),
+            "response_length": float(self.response_length),
+            "rag_documents_retrieved": float(self.rag_documents_retrieved),
+            "rag_documents_used": float(self.rag_documents_used),
+        }
+    
+    def as_params_dict(self) -> Dict[str, str]:
+        """Convert configuration fields to a params dictionary."""
+        params = {
+            "rag_enabled": str(self.rag_enabled),
+            "reranker_enabled": str(self.reranker_enabled),
+            "is_streaming": str(self.is_streaming),
+            "is_premium": str(self.is_premium),
+        }
+        if self.model_name:
+            params["model_name"] = self.model_name
+        if self.model_endpoint:
+            params["model_endpoint"] = self.model_endpoint
+        return params
+
+
+class InferenceMetricsTracker:
+    """
+    Async-compatible MLflow tracker for inference metrics.
+    
+    Uses batching and a background thread pool to avoid blocking
+    the async event loop during MLflow calls.
+    
+    Example usage in chat-handler:
+    
+        class ChatHandler:
+            def __init__(self):
+                self.mlflow_tracker = InferenceMetricsTracker(
+                    service_name="chat-handler",
+                    experiment_name="chat-inference"
+                )
+            
+            async def setup(self):
+                await self.mlflow_tracker.start()
+            
+            async def process_request(self, msg):
+                metrics = InferenceMetrics(request_id=request_id)
+                
+                # Track timing
+                start = time.time()
+                # ... do embedding ...
+                metrics.embedding_latency = time.time() - start
+                
+                # ... more processing ...
+                
+                # Log metrics (non-blocking)
+                await self.mlflow_tracker.log_inference(metrics)
+            
+            async def shutdown(self):
+                await self.mlflow_tracker.stop()
+    """
+    
+    def __init__(
+        self,
+        service_name: str,
+        experiment_name: Optional[str] = None,
+        tracking_uri: Optional[str] = None,
+        batch_size: int = 50,
+        flush_interval_seconds: float = 60.0,
+        enable_batching: bool = True,
+        max_workers: int = 2,
+    ):
+        """
+        Initialize the inference metrics tracker.
+        
+        Args:
+            service_name: Name of the service (e.g., "chat-handler")
+            experiment_name: MLflow experiment name (defaults to service_name)
+            tracking_uri: Override default tracking URI
+            batch_size: Number of metrics to batch before flushing
+            flush_interval_seconds: Maximum time between flushes
+            enable_batching: If False, log each request immediately
+            max_workers: Number of thread pool workers for MLflow calls
+        """
+        self.service_name = service_name
+        self.experiment_name = experiment_name or f"{service_name}-inference"
+        self.tracking_uri = tracking_uri
+        self.batch_size = batch_size
+        self.flush_interval = flush_interval_seconds
+        self.enable_batching = enable_batching
+        
+        self.config = MLflowConfig()
+        self._batch: List[InferenceMetrics] = []
+        self._batch_lock = asyncio.Lock()
+        self._flush_task: Optional[asyncio.Task] = None
+        self._executor = ThreadPoolExecutor(max_workers=max_workers)
+        self._running = False
+        self._client: Optional[MlflowClient] = None
+        self._experiment_id: Optional[str] = None
+        
+        # Aggregate metrics for periodic logging
+        self._aggregate_metrics: Dict[str, List[float]] = defaultdict(list)
+        self._request_count = 0
+        self._error_count = 0
+    
+    async def start(self) -> None:
+        """Start the tracker and initialize MLflow connection."""
+        if self._running:
+            return
+        
+        self._running = True
+        
+        # Initialize MLflow in thread pool to avoid blocking
+        loop = asyncio.get_event_loop()
+        await loop.run_in_executor(
+            self._executor,
+            self._init_mlflow
+        )
+        
+        if self.enable_batching:
+            self._flush_task = asyncio.create_task(self._periodic_flush())
+        
+        logger.info(
+            f"InferenceMetricsTracker started for {self.service_name} "
+            f"(experiment: {self.experiment_name})"
+        )
+    
+    def _init_mlflow(self) -> None:
+        """Initialize MLflow client and experiment (runs in thread pool)."""
+        self._client = get_mlflow_client(
+            tracking_uri=self.tracking_uri,
+            configure_global=True
+        )
+        self._experiment_id = ensure_experiment(
+            self.experiment_name,
+            tags={
+                "service": self.service_name,
+                "type": "inference-metrics",
+            }
+        )
+    
+    async def stop(self) -> None:
+        """Stop the tracker and flush remaining metrics."""
+        if not self._running:
+            return
+        
+        self._running = False
+        
+        if self._flush_task:
+            self._flush_task.cancel()
+            try:
+                await self._flush_task
+            except asyncio.CancelledError:
+                pass
+        
+        # Final flush
+        await self._flush_batch()
+        
+        self._executor.shutdown(wait=True)
+        logger.info(f"InferenceMetricsTracker stopped for {self.service_name}")
+    
+    async def log_inference(self, metrics: InferenceMetrics) -> None:
+        """
+        Log inference metrics (non-blocking).
+        
+        Args:
+            metrics: InferenceMetrics object with request data
+        """
+        if not self._running:
+            logger.warning("Tracker not running, skipping metrics")
+            return
+        
+        self._request_count += 1
+        if metrics.has_error:
+            self._error_count += 1
+        
+        # Update aggregates
+        for key, value in metrics.as_metrics_dict().items():
+            if value > 0:
+                self._aggregate_metrics[key].append(value)
+        
+        if self.enable_batching:
+            async with self._batch_lock:
+                self._batch.append(metrics)
+                if len(self._batch) >= self.batch_size:
+                    asyncio.create_task(self._flush_batch())
+        else:
+            # Immediate logging in thread pool
+            loop = asyncio.get_event_loop()
+            await loop.run_in_executor(
+                self._executor,
+                partial(self._log_single_inference, metrics)
+            )
+    
+    async def _periodic_flush(self) -> None:
+        """Periodically flush batched metrics."""
+        while self._running:
+            await asyncio.sleep(self.flush_interval)
+            await self._flush_batch()
+    
+    async def _flush_batch(self) -> None:
+        """Flush the current batch of metrics to MLflow."""
+        async with self._batch_lock:
+            if not self._batch:
+                return
+            
+            batch = self._batch
+            self._batch = []
+        
+        # Log in thread pool
+        loop = asyncio.get_event_loop()
+        await loop.run_in_executor(
+            self._executor,
+            partial(self._log_batch, batch)
+        )
+    
+    def _log_single_inference(self, metrics: InferenceMetrics) -> None:
+        """Log a single inference request to MLflow (runs in thread pool)."""
+        try:
+            with mlflow.start_run(
+                experiment_id=self._experiment_id,
+                run_name=f"inference-{metrics.request_id}",
+                tags={
+                    "service": self.service_name,
+                    "request_id": metrics.request_id,
+                    "type": "single-inference",
+                }
+            ):
+                mlflow.log_params(metrics.as_params_dict())
+                mlflow.log_metrics(metrics.as_metrics_dict())
+                
+                if metrics.user_id:
+                    mlflow.set_tag("user_id", metrics.user_id)
+                if metrics.session_id:
+                    mlflow.set_tag("session_id", metrics.session_id)
+                if metrics.has_error:
+                    mlflow.set_tag("has_error", "true")
+                    if metrics.error_message:
+                        mlflow.set_tag("error_message", metrics.error_message[:250])
+        except Exception as e:
+            logger.error(f"Failed to log inference metrics: {e}")
+    
+    def _log_batch(self, batch: List[InferenceMetrics]) -> None:
+        """Log a batch of inference metrics as aggregate statistics."""
+        if not batch:
+            return
+        
+        try:
+            # Calculate aggregates
+            aggregates = self._calculate_aggregates(batch)
+            
+            run_name = f"batch-{self.service_name}-{int(time.time())}"
+            
+            with mlflow.start_run(
+                experiment_id=self._experiment_id,
+                run_name=run_name,
+                tags={
+                    "service": self.service_name,
+                    "type": "batch-inference",
+                    "batch_size": str(len(batch)),
+                }
+            ):
+                # Log aggregate metrics
+                mlflow.log_metrics(aggregates)
+                
+                # Log batch info
+                mlflow.log_param("batch_size", len(batch))
+                mlflow.log_param("time_window_start", min(m.timestamp for m in batch))
+                mlflow.log_param("time_window_end", max(m.timestamp for m in batch))
+                
+                # Log configuration breakdown
+                rag_enabled_count = sum(1 for m in batch if m.rag_enabled)
+                streaming_count = sum(1 for m in batch if m.is_streaming)
+                premium_count = sum(1 for m in batch if m.is_premium)
+                error_count = sum(1 for m in batch if m.has_error)
+                
+                mlflow.log_metrics({
+                    "rag_enabled_pct": rag_enabled_count / len(batch) * 100,
+                    "streaming_pct": streaming_count / len(batch) * 100,
+                    "premium_pct": premium_count / len(batch) * 100,
+                    "error_rate": error_count / len(batch) * 100,
+                })
+                
+                # Log model distribution
+                model_counts: Dict[str, int] = defaultdict(int)
+                for m in batch:
+                    if m.model_name:
+                        model_counts[m.model_name] += 1
+                
+                if model_counts:
+                    mlflow.log_dict(
+                        {"models": dict(model_counts)},
+                        "model_distribution.json"
+                    )
+            
+            logger.info(f"Logged batch of {len(batch)} inference metrics")
+            
+        except Exception as e:
+            logger.error(f"Failed to log batch metrics: {e}")
+    
+    def _calculate_aggregates(
+        self,
+        batch: List[InferenceMetrics]
+    ) -> Dict[str, float]:
+        """Calculate aggregate statistics from a batch of metrics."""
+        import statistics
+        
+        aggregates = {}
+        
+        # Collect all numeric metrics
+        metric_values: Dict[str, List[float]] = defaultdict(list)
+        for m in batch:
+            for key, value in m.as_metrics_dict().items():
+                if value > 0:
+                    metric_values[key].append(value)
+        
+        # Calculate statistics for each metric
+        for key, values in metric_values.items():
+            if not values:
+                continue
+            
+            aggregates[f"{key}_mean"] = statistics.mean(values)
+            aggregates[f"{key}_min"] = min(values)
+            aggregates[f"{key}_max"] = max(values)
+            
+            if len(values) >= 2:
+                aggregates[f"{key}_p50"] = statistics.median(values)
+                aggregates[f"{key}_stdev"] = statistics.stdev(values)
+            
+            if len(values) >= 4:
+                sorted_vals = sorted(values)
+                p95_idx = int(len(sorted_vals) * 0.95)
+                p99_idx = int(len(sorted_vals) * 0.99)
+                aggregates[f"{key}_p95"] = sorted_vals[p95_idx]
+                aggregates[f"{key}_p99"] = sorted_vals[p99_idx]
+        
+        # Add counts
+        aggregates["total_requests"] = float(len(batch))
+        
+        return aggregates
+    
+    def get_stats(self) -> Dict[str, Any]:
+        """Get current tracker statistics."""
+        return {
+            "service_name": self.service_name,
+            "experiment_name": self.experiment_name,
+            "running": self._running,
+            "total_requests": self._request_count,
+            "error_count": self._error_count,
+            "pending_batch_size": len(self._batch),
+            "aggregate_metrics_count": len(self._aggregate_metrics),
+        }
diff --git a/mlflow_utils/kfp_components.py b/mlflow_utils/kfp_components.py
new file mode 100644
index 0000000..f3cdf18
--- /dev/null
+++ b/mlflow_utils/kfp_components.py
@@ -0,0 +1,513 @@
+"""
+Kubeflow Pipeline Components with MLflow Tracking
+
+Provides reusable KFP components that integrate MLflow experiment
+tracking into Kubeflow Pipelines. These components can be used
+directly in pipelines or as wrappers around existing pipeline steps.
+
+Usage in a Kubeflow Pipeline:
+
+    from mlflow_utils.kfp_components import (
+        create_mlflow_run,
+        log_metrics_component,
+        log_model_artifact,
+        end_mlflow_run,
+    )
+
+    @dsl.pipeline(name="my-pipeline")
+    def my_pipeline():
+        # Start MLflow run
+        run_info = create_mlflow_run(
+            experiment_name="my-experiment",
+            run_name="training-run-1"
+        )
+        
+        # ... your pipeline steps ...
+        
+        # Log metrics
+        log_step = log_metrics_component(
+            run_id=run_info.outputs["run_id"],
+            metrics={"accuracy": 0.95, "loss": 0.05}
+        )
+        
+        # End run
+        end_mlflow_run(run_id=run_info.outputs["run_id"])
+"""
+
+from kfp import dsl
+from typing import Dict, Any, List, Optional, NamedTuple
+
+
+# MLflow component image with all required dependencies
+MLFLOW_IMAGE = "python:3.13-slim"
+MLFLOW_PACKAGES = [
+    "mlflow>=2.10.0",
+    "boto3",  # For S3 artifact storage if needed
+    "psycopg2-binary",  # For PostgreSQL backend
+]
+
+
+@dsl.component(
+    base_image=MLFLOW_IMAGE,
+    packages_to_install=MLFLOW_PACKAGES
+)
+def create_mlflow_run(
+    experiment_name: str,
+    run_name: str,
+    mlflow_tracking_uri: str = "http://mlflow.mlflow.svc.cluster.local:80",
+    tags: Dict[str, str] = None,
+    params: Dict[str, str] = None,
+) -> NamedTuple('RunInfo', [('run_id', str), ('experiment_id', str), ('artifact_uri', str)]):
+    """
+    Create a new MLflow run for the pipeline.
+    
+    This should be called at the start of a pipeline to initialize
+    tracking. The returned run_id should be passed to subsequent
+    components for logging.
+    
+    Args:
+        experiment_name: Name of the MLflow experiment
+        run_name: Name for this specific run
+        mlflow_tracking_uri: MLflow tracking server URI
+        tags: Optional tags to add to the run
+        params: Optional parameters to log
+        
+    Returns:
+        NamedTuple with run_id, experiment_id, and artifact_uri
+    """
+    import os
+    import mlflow
+    from mlflow.tracking import MlflowClient
+    from collections import namedtuple
+    
+    # Set tracking URI
+    mlflow.set_tracking_uri(mlflow_tracking_uri)
+    
+    client = MlflowClient()
+    
+    # Get or create experiment
+    experiment = client.get_experiment_by_name(experiment_name)
+    if experiment is None:
+        experiment_id = client.create_experiment(
+            name=experiment_name,
+            artifact_location=f"/mlflow/artifacts/{experiment_name}"
+        )
+    else:
+        experiment_id = experiment.experiment_id
+    
+    # Create default tags
+    default_tags = {
+        "pipeline.type": "kubeflow",
+        "kfp.run_id": os.environ.get("KFP_RUN_ID", "unknown"),
+        "kfp.pod_name": os.environ.get("HOSTNAME", "unknown"),
+    }
+    if tags:
+        default_tags.update(tags)
+    
+    # Start run
+    run = mlflow.start_run(
+        experiment_id=experiment_id,
+        run_name=run_name,
+        tags=default_tags,
+    )
+    
+    # Log initial params
+    if params:
+        mlflow.log_params(params)
+    
+    run_id = run.info.run_id
+    artifact_uri = run.info.artifact_uri
+    
+    # End run (KFP components are isolated, we'll resume in other components)
+    mlflow.end_run()
+    
+    RunInfo = namedtuple('RunInfo', ['run_id', 'experiment_id', 'artifact_uri'])
+    return RunInfo(run_id, experiment_id, artifact_uri)
+
+
+@dsl.component(
+    base_image=MLFLOW_IMAGE,
+    packages_to_install=MLFLOW_PACKAGES
+)
+def log_params_component(
+    run_id: str,
+    params: Dict[str, str],
+    mlflow_tracking_uri: str = "http://mlflow.mlflow.svc.cluster.local:80",
+) -> str:
+    """
+    Log parameters to an existing MLflow run.
+    
+    Args:
+        run_id: The MLflow run ID to log to
+        params: Dictionary of parameters to log
+        mlflow_tracking_uri: MLflow tracking server URI
+        
+    Returns:
+        The run_id for chaining
+    """
+    import mlflow
+    from mlflow.tracking import MlflowClient
+    
+    mlflow.set_tracking_uri(mlflow_tracking_uri)
+    client = MlflowClient()
+    
+    for key, value in params.items():
+        client.log_param(run_id, key, str(value)[:500])
+    
+    return run_id
+
+
+@dsl.component(
+    base_image=MLFLOW_IMAGE,
+    packages_to_install=MLFLOW_PACKAGES
+)
+def log_metrics_component(
+    run_id: str,
+    metrics: Dict[str, float],
+    step: int = 0,
+    mlflow_tracking_uri: str = "http://mlflow.mlflow.svc.cluster.local:80",
+) -> str:
+    """
+    Log metrics to an existing MLflow run.
+    
+    Args:
+        run_id: The MLflow run ID to log to
+        metrics: Dictionary of metrics to log
+        step: Step number for time-series metrics
+        mlflow_tracking_uri: MLflow tracking server URI
+        
+    Returns:
+        The run_id for chaining
+    """
+    import mlflow
+    from mlflow.tracking import MlflowClient
+    
+    mlflow.set_tracking_uri(mlflow_tracking_uri)
+    client = MlflowClient()
+    
+    for key, value in metrics.items():
+        client.log_metric(run_id, key, float(value), step=step)
+    
+    return run_id
+
+
+@dsl.component(
+    base_image=MLFLOW_IMAGE,
+    packages_to_install=MLFLOW_PACKAGES
+)
+def log_artifact_component(
+    run_id: str,
+    artifact_path: str,
+    artifact_name: str = "",
+    mlflow_tracking_uri: str = "http://mlflow.mlflow.svc.cluster.local:80",
+) -> str:
+    """
+    Log an artifact file to an existing MLflow run.
+    
+    Args:
+        run_id: The MLflow run ID to log to
+        artifact_path: Path to the artifact file
+        artifact_name: Optional destination name in artifact store
+        mlflow_tracking_uri: MLflow tracking server URI
+        
+    Returns:
+        The run_id for chaining
+    """
+    import mlflow
+    from mlflow.tracking import MlflowClient
+    
+    mlflow.set_tracking_uri(mlflow_tracking_uri)
+    client = MlflowClient()
+    
+    client.log_artifact(run_id, artifact_path, artifact_name or None)
+    
+    return run_id
+
+
+@dsl.component(
+    base_image=MLFLOW_IMAGE,
+    packages_to_install=MLFLOW_PACKAGES
+)
+def log_dict_artifact(
+    run_id: str,
+    data: Dict[str, Any],
+    filename: str,
+    mlflow_tracking_uri: str = "http://mlflow.mlflow.svc.cluster.local:80",
+) -> str:
+    """
+    Log a dictionary as a JSON artifact.
+    
+    Args:
+        run_id: The MLflow run ID to log to
+        data: Dictionary to save as JSON
+        filename: Name for the JSON file
+        mlflow_tracking_uri: MLflow tracking server URI
+        
+    Returns:
+        The run_id for chaining
+    """
+    import json
+    import tempfile
+    import mlflow
+    from mlflow.tracking import MlflowClient
+    from pathlib import Path
+    
+    mlflow.set_tracking_uri(mlflow_tracking_uri)
+    client = MlflowClient()
+    
+    # Ensure .json extension
+    if not filename.endswith('.json'):
+        filename += '.json'
+    
+    # Write to temp file and log
+    with tempfile.TemporaryDirectory() as tmpdir:
+        filepath = Path(tmpdir) / filename
+        with open(filepath, 'w') as f:
+            json.dump(data, f, indent=2)
+        client.log_artifact(run_id, str(filepath))
+    
+    return run_id
+
+
+@dsl.component(
+    base_image=MLFLOW_IMAGE,
+    packages_to_install=MLFLOW_PACKAGES
+)
+def end_mlflow_run(
+    run_id: str,
+    status: str = "FINISHED",
+    mlflow_tracking_uri: str = "http://mlflow.mlflow.svc.cluster.local:80",
+) -> str:
+    """
+    End an MLflow run with the specified status.
+    
+    Args:
+        run_id: The MLflow run ID to end
+        status: Run status (FINISHED, FAILED, KILLED)
+        mlflow_tracking_uri: MLflow tracking server URI
+        
+    Returns:
+        The run_id
+    """
+    import mlflow
+    from mlflow.tracking import MlflowClient
+    from mlflow.entities import RunStatus
+    
+    mlflow.set_tracking_uri(mlflow_tracking_uri)
+    client = MlflowClient()
+    
+    status_map = {
+        "FINISHED": RunStatus.FINISHED,
+        "FAILED": RunStatus.FAILED,
+        "KILLED": RunStatus.KILLED,
+    }
+    
+    run_status = status_map.get(status.upper(), RunStatus.FINISHED)
+    client.set_terminated(run_id, status=run_status)
+    
+    return run_id
+
+
+@dsl.component(
+    base_image=MLFLOW_IMAGE,
+    packages_to_install=MLFLOW_PACKAGES + ["httpx"]
+)
+def log_training_metrics(
+    run_id: str,
+    model_type: str,
+    training_config: Dict[str, Any],
+    final_metrics: Dict[str, float],
+    model_path: str = "",
+    mlflow_tracking_uri: str = "http://mlflow.mlflow.svc.cluster.local:80",
+) -> str:
+    """
+    Log comprehensive training metrics for ML models.
+    
+    Designed for use with QLoRA training, voice training, and other
+    ML training pipelines in the llm-workflows repository.
+    
+    Args:
+        run_id: The MLflow run ID to log to
+        model_type: Type of model (llm, stt, tts, embeddings)
+        training_config: Training configuration dict
+        final_metrics: Final training metrics
+        model_path: Path to saved model (if applicable)
+        mlflow_tracking_uri: MLflow tracking server URI
+        
+    Returns:
+        The run_id for chaining
+    """
+    import json
+    import tempfile
+    import mlflow
+    from mlflow.tracking import MlflowClient
+    from pathlib import Path
+    
+    mlflow.set_tracking_uri(mlflow_tracking_uri)
+    client = MlflowClient()
+    
+    # Log training config as params
+    flat_config = {}
+    for key, value in training_config.items():
+        if isinstance(value, (dict, list)):
+            flat_config[f"config.{key}"] = json.dumps(value)[:500]
+        else:
+            flat_config[f"config.{key}"] = str(value)[:500]
+    
+    for key, value in flat_config.items():
+        client.log_param(run_id, key, value)
+    
+    # Log model type tag
+    client.set_tag(run_id, "model.type", model_type)
+    
+    # Log metrics
+    for key, value in final_metrics.items():
+        client.log_metric(run_id, key, float(value))
+    
+    # Log full config as artifact
+    with tempfile.TemporaryDirectory() as tmpdir:
+        config_path = Path(tmpdir) / "training_config.json"
+        with open(config_path, 'w') as f:
+            json.dump(training_config, f, indent=2)
+        client.log_artifact(run_id, str(config_path))
+    
+    # Log model path if provided
+    if model_path:
+        client.log_param(run_id, "model.path", model_path)
+        client.set_tag(run_id, "model.saved", "true")
+    
+    return run_id
+
+
+@dsl.component(
+    base_image=MLFLOW_IMAGE,
+    packages_to_install=MLFLOW_PACKAGES
+)
+def log_document_ingestion_metrics(
+    run_id: str,
+    source_url: str,
+    collection_name: str,
+    chunks_created: int,
+    documents_processed: int,
+    processing_time_seconds: float,
+    embeddings_model: str = "bge-small-en-v1.5",
+    chunk_size: int = 500,
+    chunk_overlap: int = 50,
+    mlflow_tracking_uri: str = "http://mlflow.mlflow.svc.cluster.local:80",
+) -> str:
+    """
+    Log document ingestion pipeline metrics.
+    
+    Designed for use with the document_ingestion_pipeline.
+    
+    Args:
+        run_id: The MLflow run ID to log to
+        source_url: URL of the source document
+        collection_name: Milvus collection name
+        chunks_created: Number of chunks created
+        documents_processed: Number of documents processed
+        processing_time_seconds: Total processing time
+        embeddings_model: Embeddings model used
+        chunk_size: Chunk size in tokens
+        chunk_overlap: Chunk overlap in tokens
+        mlflow_tracking_uri: MLflow tracking server URI
+        
+    Returns:
+        The run_id for chaining
+    """
+    import mlflow
+    from mlflow.tracking import MlflowClient
+    
+    mlflow.set_tracking_uri(mlflow_tracking_uri)
+    client = MlflowClient()
+    
+    # Log params
+    params = {
+        "source_url": source_url[:500],
+        "collection_name": collection_name,
+        "embeddings_model": embeddings_model,
+        "chunk_size": str(chunk_size),
+        "chunk_overlap": str(chunk_overlap),
+    }
+    for key, value in params.items():
+        client.log_param(run_id, key, value)
+    
+    # Log metrics
+    metrics = {
+        "chunks_created": chunks_created,
+        "documents_processed": documents_processed,
+        "processing_time_seconds": processing_time_seconds,
+        "chunks_per_second": chunks_created / processing_time_seconds if processing_time_seconds > 0 else 0,
+    }
+    for key, value in metrics.items():
+        client.log_metric(run_id, key, float(value))
+    
+    # Set pipeline type tag
+    client.set_tag(run_id, "pipeline.type", "document-ingestion")
+    client.set_tag(run_id, "milvus.collection", collection_name)
+    
+    return run_id
+
+
+@dsl.component(
+    base_image=MLFLOW_IMAGE,
+    packages_to_install=MLFLOW_PACKAGES
+)
+def log_evaluation_results(
+    run_id: str,
+    model_name: str,
+    dataset_name: str,
+    metrics: Dict[str, float],
+    sample_results: List[Dict[str, Any]] = None,
+    mlflow_tracking_uri: str = "http://mlflow.mlflow.svc.cluster.local:80",
+) -> str:
+    """
+    Log model evaluation results.
+    
+    Designed for use with the evaluation_pipeline.
+    
+    Args:
+        run_id: The MLflow run ID to log to
+        model_name: Name of the evaluated model
+        dataset_name: Name of the evaluation dataset
+        metrics: Evaluation metrics (accuracy, etc.)
+        sample_results: Optional sample predictions
+        mlflow_tracking_uri: MLflow tracking server URI
+        
+    Returns:
+        The run_id for chaining
+    """
+    import json
+    import tempfile
+    import mlflow
+    from mlflow.tracking import MlflowClient
+    from pathlib import Path
+    
+    mlflow.set_tracking_uri(mlflow_tracking_uri)
+    client = MlflowClient()
+    
+    # Log params
+    client.log_param(run_id, "eval.model_name", model_name)
+    client.log_param(run_id, "eval.dataset", dataset_name)
+    
+    # Log metrics
+    for key, value in metrics.items():
+        client.log_metric(run_id, f"eval.{key}", float(value))
+    
+    # Log sample results as artifact
+    if sample_results:
+        with tempfile.TemporaryDirectory() as tmpdir:
+            results_path = Path(tmpdir) / "evaluation_results.json"
+            with open(results_path, 'w') as f:
+                json.dump(sample_results, f, indent=2)
+            client.log_artifact(run_id, str(results_path))
+    
+    # Set tags
+    client.set_tag(run_id, "pipeline.type", "evaluation")
+    client.set_tag(run_id, "model.name", model_name)
+    
+    # Determine if passed
+    passed = metrics.get("pass", metrics.get("accuracy", 0) >= 0.7)
+    client.set_tag(run_id, "eval.passed", str(passed))
+    
+    return run_id
diff --git a/mlflow_utils/model_registry.py b/mlflow_utils/model_registry.py
new file mode 100644
index 0000000..ccc7674
--- /dev/null
+++ b/mlflow_utils/model_registry.py
@@ -0,0 +1,545 @@
+"""
+MLflow Model Registry Integration for KServe
+
+Provides utilities for registering trained models in MLflow Model Registry
+with metadata needed for deployment to KServe InferenceServices.
+
+This module bridges the gap between Kubeflow training pipelines and
+KServe model serving by:
+1. Registering models with proper versioning
+2. Adding KServe-specific metadata (runtime, protocol, resources)
+3. Managing model stage transitions (Staging → Production)
+4. Generating KServe InferenceService manifests from registered models
+
+Usage:
+    from mlflow_utils.model_registry import (
+        register_model_for_kserve,
+        promote_model_to_production,
+        generate_kserve_manifest,
+    )
+    
+    # Register a new model version
+    model_version = register_model_for_kserve(
+        model_name="whisper-finetuned",
+        model_uri="s3://models/whisper-v2",
+        model_type="stt",
+        kserve_config={
+            "runtime": "kserve-huggingface",
+            "container_image": "ghcr.io/my-org/whisper:v2",
+        }
+    )
+    
+    # Generate KServe manifest for deployment
+    manifest = generate_kserve_manifest(
+        model_name="whisper-finetuned",
+        model_version=model_version.version,
+    )
+"""
+
+import os
+import json
+import yaml
+import logging
+from typing import Optional, Dict, Any, List
+from dataclasses import dataclass, field
+
+import mlflow
+from mlflow.tracking import MlflowClient
+from mlflow.entities.model_registry import ModelVersion
+
+from .client import get_mlflow_client, MLflowConfig
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class KServeConfig:
+    """Configuration for KServe deployment."""
+    
+    # Runtime/container configuration
+    runtime: str = "kserve-huggingface"  # kserve-huggingface, kserve-custom, etc.
+    container_image: Optional[str] = None
+    container_port: int = 8080
+    
+    # Protocol configuration
+    protocol: str = "v2"  # v1, v2, grpc
+    
+    # Resource requests/limits
+    cpu_request: str = "1"
+    cpu_limit: str = "4"
+    memory_request: str = "4Gi"
+    memory_limit: str = "16Gi"
+    gpu_count: int = 0
+    gpu_type: str = "nvidia.com/gpu"  # or amd.com/gpu for ROCm
+    
+    # Storage configuration
+    storage_uri: Optional[str] = None  # s3://, pvc://, gs://
+    
+    # Scaling configuration
+    min_replicas: int = 1
+    max_replicas: int = 1
+    scale_target: int = 10  # Target concurrent requests for scaling
+    
+    # Serving configuration
+    timeout_seconds: int = 300
+    batch_size: int = 1
+    
+    # Additional environment variables
+    env_vars: Dict[str, str] = field(default_factory=dict)
+    
+    def as_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for MLflow tags."""
+        return {
+            "kserve.runtime": self.runtime,
+            "kserve.protocol": self.protocol,
+            "kserve.cpu_request": self.cpu_request,
+            "kserve.memory_request": self.memory_request,
+            "kserve.gpu_count": str(self.gpu_count),
+            "kserve.min_replicas": str(self.min_replicas),
+            "kserve.max_replicas": str(self.max_replicas),
+            "kserve.storage_uri": self.storage_uri or "",
+            "kserve.container_image": self.container_image or "",
+        }
+
+
+# Pre-configured KServe configurations for common model types
+KSERVE_PRESETS: Dict[str, KServeConfig] = {
+    "llm": KServeConfig(
+        runtime="kserve-huggingface",
+        cpu_request="2",
+        cpu_limit="8",
+        memory_request="16Gi",
+        memory_limit="64Gi",
+        gpu_count=1,
+        timeout_seconds=600,
+    ),
+    "stt": KServeConfig(
+        runtime="kserve-custom",
+        cpu_request="2",
+        cpu_limit="4",
+        memory_request="8Gi",
+        memory_limit="16Gi",
+        gpu_count=1,
+        timeout_seconds=120,
+    ),
+    "tts": KServeConfig(
+        runtime="kserve-custom",
+        cpu_request="2",
+        cpu_limit="4",
+        memory_request="8Gi",
+        memory_limit="16Gi",
+        gpu_count=1,
+        timeout_seconds=60,
+    ),
+    "embeddings": KServeConfig(
+        runtime="kserve-huggingface",
+        cpu_request="1",
+        cpu_limit="4",
+        memory_request="4Gi",
+        memory_limit="16Gi",
+        gpu_count=0,
+        timeout_seconds=30,
+        batch_size=32,
+    ),
+    "reranker": KServeConfig(
+        runtime="kserve-huggingface",
+        cpu_request="1",
+        cpu_limit="4",
+        memory_request="4Gi",
+        memory_limit="16Gi",
+        gpu_count=0,
+        timeout_seconds=30,
+    ),
+}
+
+
+def register_model_for_kserve(
+    model_name: str,
+    model_uri: str,
+    model_type: str,
+    run_id: Optional[str] = None,
+    description: Optional[str] = None,
+    kserve_config: Optional[KServeConfig] = None,
+    tags: Optional[Dict[str, str]] = None,
+    tracking_uri: Optional[str] = None,
+) -> ModelVersion:
+    """
+    Register a model in MLflow Model Registry with KServe metadata.
+    
+    Args:
+        model_name: Name for the registered model
+        model_uri: URI to model artifacts (runs:/run_id/path, s3://, pvc://)
+        model_type: Type of model (llm, stt, tts, embeddings, reranker)
+        run_id: Optional MLflow run ID to associate with
+        description: Description of the model version
+        kserve_config: KServe deployment configuration
+        tags: Additional tags for the model version
+        tracking_uri: Override default tracking URI
+        
+    Returns:
+        The created ModelVersion object
+    """
+    client = get_mlflow_client(tracking_uri=tracking_uri)
+    
+    # Get or use preset KServe config
+    if kserve_config is None:
+        kserve_config = KSERVE_PRESETS.get(model_type, KServeConfig())
+    
+    # Ensure registered model exists
+    try:
+        client.get_registered_model(model_name)
+    except mlflow.exceptions.MlflowException:
+        client.create_registered_model(
+            name=model_name,
+            description=f"{model_type.upper()} model for KServe deployment",
+            tags={
+                "model.type": model_type,
+                "deployment.target": "kserve",
+            }
+        )
+        logger.info(f"Created registered model: {model_name}")
+    
+    # Create model version
+    model_version = client.create_model_version(
+        name=model_name,
+        source=model_uri,
+        run_id=run_id,
+        description=description or f"Version from {model_uri}",
+        tags={
+            **(tags or {}),
+            "model.type": model_type,
+            **kserve_config.as_dict(),
+        }
+    )
+    
+    logger.info(
+        f"Registered model version {model_version.version} "
+        f"for {model_name} (type: {model_type})"
+    )
+    
+    return model_version
+
+
+def promote_model_to_stage(
+    model_name: str,
+    version: int,
+    stage: str = "Staging",
+    archive_existing: bool = True,
+    tracking_uri: Optional[str] = None,
+) -> ModelVersion:
+    """
+    Promote a model version to a new stage.
+    
+    Args:
+        model_name: Name of the registered model
+        version: Version number to promote
+        stage: Target stage (Staging, Production, Archived)
+        archive_existing: If True, archive existing versions in target stage
+        tracking_uri: Override default tracking URI
+        
+    Returns:
+        The updated ModelVersion
+    """
+    client = get_mlflow_client(tracking_uri=tracking_uri)
+    
+    # Transition to new stage
+    model_version = client.transition_model_version_stage(
+        name=model_name,
+        version=str(version),
+        stage=stage,
+        archive_existing_versions=archive_existing,
+    )
+    
+    logger.info(f"Promoted {model_name} v{version} to {stage}")
+    
+    return model_version
+
+
+def promote_model_to_production(
+    model_name: str,
+    version: int,
+    tracking_uri: Optional[str] = None,
+) -> ModelVersion:
+    """
+    Promote a model version directly to Production.
+    
+    Args:
+        model_name: Name of the registered model
+        version: Version number to promote
+        tracking_uri: Override default tracking URI
+        
+    Returns:
+        The updated ModelVersion
+    """
+    return promote_model_to_stage(
+        model_name=model_name,
+        version=version,
+        stage="Production",
+        archive_existing=True,
+        tracking_uri=tracking_uri,
+    )
+
+
+def get_production_model(
+    model_name: str,
+    tracking_uri: Optional[str] = None,
+) -> Optional[ModelVersion]:
+    """
+    Get the current Production model version.
+    
+    Args:
+        model_name: Name of the registered model
+        tracking_uri: Override default tracking URI
+        
+    Returns:
+        The Production ModelVersion, or None if none exists
+    """
+    client = get_mlflow_client(tracking_uri=tracking_uri)
+    
+    versions = client.get_latest_versions(model_name, stages=["Production"])
+    
+    return versions[0] if versions else None
+
+
+def get_model_kserve_config(
+    model_name: str,
+    version: Optional[int] = None,
+    tracking_uri: Optional[str] = None,
+) -> KServeConfig:
+    """
+    Get KServe configuration from a registered model version.
+    
+    Args:
+        model_name: Name of the registered model
+        version: Version number (uses Production if not specified)
+        tracking_uri: Override default tracking URI
+        
+    Returns:
+        KServeConfig populated from model tags
+    """
+    client = get_mlflow_client(tracking_uri=tracking_uri)
+    
+    if version:
+        model_version = client.get_model_version(model_name, str(version))
+    else:
+        prod_version = get_production_model(model_name, tracking_uri)
+        if not prod_version:
+            raise ValueError(f"No Production version for {model_name}")
+        model_version = prod_version
+    
+    tags = model_version.tags
+    
+    return KServeConfig(
+        runtime=tags.get("kserve.runtime", "kserve-huggingface"),
+        protocol=tags.get("kserve.protocol", "v2"),
+        cpu_request=tags.get("kserve.cpu_request", "1"),
+        memory_request=tags.get("kserve.memory_request", "4Gi"),
+        gpu_count=int(tags.get("kserve.gpu_count", "0")),
+        min_replicas=int(tags.get("kserve.min_replicas", "1")),
+        max_replicas=int(tags.get("kserve.max_replicas", "1")),
+        storage_uri=tags.get("kserve.storage_uri") or None,
+        container_image=tags.get("kserve.container_image") or None,
+    )
+
+
+def generate_kserve_manifest(
+    model_name: str,
+    version: Optional[int] = None,
+    namespace: str = "ai-ml",
+    service_name: Optional[str] = None,
+    extra_annotations: Optional[Dict[str, str]] = None,
+    tracking_uri: Optional[str] = None,
+) -> Dict[str, Any]:
+    """
+    Generate a KServe InferenceService manifest from a registered model.
+    
+    Args:
+        model_name: Name of the registered model
+        version: Version number (uses Production if not specified)
+        namespace: Kubernetes namespace for deployment
+        service_name: Name for the InferenceService (defaults to model_name)
+        extra_annotations: Additional annotations for the service
+        tracking_uri: Override default tracking URI
+        
+    Returns:
+        KServe InferenceService manifest as a dictionary
+    """
+    client = get_mlflow_client(tracking_uri=tracking_uri)
+    
+    # Get model version
+    if version:
+        model_version = client.get_model_version(model_name, str(version))
+    else:
+        prod_version = get_production_model(model_name, tracking_uri)
+        if not prod_version:
+            raise ValueError(f"No Production version for {model_name}")
+        model_version = prod_version
+        version = int(model_version.version)
+    
+    # Get KServe config
+    config = get_model_kserve_config(model_name, version, tracking_uri)
+    model_type = model_version.tags.get("model.type", "custom")
+    
+    svc_name = service_name or model_name.lower().replace("_", "-")
+    
+    # Build manifest
+    manifest = {
+        "apiVersion": "serving.kserve.io/v1beta1",
+        "kind": "InferenceService",
+        "metadata": {
+            "name": svc_name,
+            "namespace": namespace,
+            "labels": {
+                "mlflow.model": model_name,
+                "mlflow.version": str(version),
+                "model.type": model_type,
+            },
+            "annotations": {
+                "mlflow.tracking_uri": get_mlflow_client().tracking_uri,
+                "mlflow.run_id": model_version.run_id or "",
+                **(extra_annotations or {}),
+            },
+        },
+        "spec": {
+            "predictor": {
+                "minReplicas": config.min_replicas,
+                "maxReplicas": config.max_replicas,
+                "scaleTarget": config.scale_target,
+                "timeout": config.timeout_seconds,
+            },
+        },
+    }
+    
+    # Configure predictor based on runtime
+    predictor = manifest["spec"]["predictor"]
+    
+    if config.container_image:
+        # Custom container
+        predictor["containers"] = [{
+            "name": "predictor",
+            "image": config.container_image,
+            "ports": [{"containerPort": config.container_port, "protocol": "TCP"}],
+            "resources": {
+                "requests": {
+                    "cpu": config.cpu_request,
+                    "memory": config.memory_request,
+                },
+                "limits": {
+                    "cpu": config.cpu_limit,
+                    "memory": config.memory_limit,
+                },
+            },
+            "env": [
+                {"name": k, "value": v}
+                for k, v in config.env_vars.items()
+            ],
+        }]
+        
+        # Add GPU if needed
+        if config.gpu_count > 0:
+            predictor["containers"][0]["resources"]["limits"][config.gpu_type] = str(config.gpu_count)
+            predictor["containers"][0]["resources"]["requests"][config.gpu_type] = str(config.gpu_count)
+            
+    else:
+        # Standard KServe runtime
+        storage_uri = config.storage_uri or model_version.source
+        
+        predictor["model"] = {
+            "modelFormat": {"name": "huggingface"},
+            "protocolVersion": config.protocol,
+            "storageUri": storage_uri,
+            "resources": {
+                "requests": {
+                    "cpu": config.cpu_request,
+                    "memory": config.memory_request,
+                },
+                "limits": {
+                    "cpu": config.cpu_limit,
+                    "memory": config.memory_limit,
+                },
+            },
+        }
+        
+        if config.gpu_count > 0:
+            predictor["model"]["resources"]["limits"][config.gpu_type] = str(config.gpu_count)
+            predictor["model"]["resources"]["requests"][config.gpu_type] = str(config.gpu_count)
+    
+    return manifest
+
+
+def generate_kserve_yaml(
+    model_name: str,
+    version: Optional[int] = None,
+    namespace: str = "ai-ml",
+    output_path: Optional[str] = None,
+    tracking_uri: Optional[str] = None,
+) -> str:
+    """
+    Generate a KServe InferenceService manifest as YAML.
+    
+    Args:
+        model_name: Name of the registered model
+        version: Version number (uses Production if not specified)
+        namespace: Kubernetes namespace
+        output_path: If provided, write YAML to this path
+        tracking_uri: Override default tracking URI
+        
+    Returns:
+        YAML string of the manifest
+    """
+    manifest = generate_kserve_manifest(
+        model_name=model_name,
+        version=version,
+        namespace=namespace,
+        tracking_uri=tracking_uri,
+    )
+    
+    yaml_str = yaml.dump(manifest, default_flow_style=False, sort_keys=False)
+    
+    if output_path:
+        with open(output_path, 'w') as f:
+            f.write(yaml_str)
+        logger.info(f"Wrote KServe manifest to {output_path}")
+    
+    return yaml_str
+
+
+def list_model_versions(
+    model_name: str,
+    stages: Optional[List[str]] = None,
+    tracking_uri: Optional[str] = None,
+) -> List[Dict[str, Any]]:
+    """
+    List all versions of a registered model.
+    
+    Args:
+        model_name: Name of the registered model
+        stages: Filter by stages (None for all)
+        tracking_uri: Override default tracking URI
+        
+    Returns:
+        List of model version info dictionaries
+    """
+    client = get_mlflow_client(tracking_uri=tracking_uri)
+    
+    if stages:
+        versions = client.get_latest_versions(model_name, stages=stages)
+    else:
+        # Get all versions
+        versions = []
+        for mv in client.search_model_versions(f"name='{model_name}'"):
+            versions.append(mv)
+    
+    return [
+        {
+            "version": mv.version,
+            "stage": mv.current_stage,
+            "source": mv.source,
+            "run_id": mv.run_id,
+            "description": mv.description,
+            "tags": mv.tags,
+            "creation_timestamp": mv.creation_timestamp,
+            "last_updated_timestamp": mv.last_updated_timestamp,
+        }
+        for mv in versions
+    ]
diff --git a/mlflow_utils/tracker.py b/mlflow_utils/tracker.py
new file mode 100644
index 0000000..a92ae6b
--- /dev/null
+++ b/mlflow_utils/tracker.py
@@ -0,0 +1,395 @@
+"""
+MLflow Tracker for Kubeflow Pipelines
+
+Provides a high-level interface for logging experiments, parameters,
+metrics, and artifacts from Kubeflow Pipeline components.
+"""
+
+import os
+import json
+import time
+import logging
+from pathlib import Path
+from typing import Optional, Dict, Any, List, Union
+from contextlib import contextmanager
+from dataclasses import dataclass, field
+
+import mlflow
+from mlflow.tracking import MlflowClient
+
+from .client import get_mlflow_client, ensure_experiment, MLflowConfig
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class PipelineMetadata:
+    """Metadata about the Kubeflow Pipeline run."""
+    pipeline_name: str
+    run_id: str
+    run_name: Optional[str] = None
+    component_name: Optional[str] = None
+    namespace: str = "ai-ml"
+    
+    # KFP-specific metadata (populated from environment if available)
+    kfp_run_id: Optional[str] = field(
+        default_factory=lambda: os.environ.get("KFP_RUN_ID")
+    )
+    kfp_pod_name: Optional[str] = field(
+        default_factory=lambda: os.environ.get("KFP_POD_NAME")
+    )
+    
+    def as_tags(self) -> Dict[str, str]:
+        """Convert metadata to MLflow tags."""
+        tags = {
+            "pipeline.name": self.pipeline_name,
+            "pipeline.run_id": self.run_id,
+            "pipeline.namespace": self.namespace,
+        }
+        if self.run_name:
+            tags["pipeline.run_name"] = self.run_name
+        if self.component_name:
+            tags["pipeline.component"] = self.component_name
+        if self.kfp_run_id:
+            tags["kfp.run_id"] = self.kfp_run_id
+        if self.kfp_pod_name:
+            tags["kfp.pod_name"] = self.kfp_pod_name
+        return tags
+
+
+class MLflowTracker:
+    """
+    MLflow experiment tracker for Kubeflow Pipeline components.
+    
+    Example usage in a KFP component:
+    
+        from mlflow_utils import MLflowTracker
+        
+        tracker = MLflowTracker(
+            experiment_name="document-ingestion",
+            run_name="batch-ingestion-2024-01"
+        )
+        
+        with tracker.start_run() as run:
+            tracker.log_params({
+                "chunk_size": 500,
+                "overlap": 50,
+                "embeddings_model": "bge-small-en-v1.5"
+            })
+            
+            # ... do work ...
+            
+            tracker.log_metrics({
+                "documents_processed": 100,
+                "chunks_created": 2500,
+                "processing_time_seconds": 120.5
+            })
+            
+            tracker.log_artifact("/path/to/output.json")
+    """
+    
+    def __init__(
+        self,
+        experiment_name: str,
+        run_name: Optional[str] = None,
+        pipeline_metadata: Optional[PipelineMetadata] = None,
+        tags: Optional[Dict[str, str]] = None,
+        tracking_uri: Optional[str] = None,
+    ):
+        """
+        Initialize the MLflow tracker.
+        
+        Args:
+            experiment_name: Name of the MLflow experiment
+            run_name: Optional name for this run
+            pipeline_metadata: Metadata about the KFP pipeline
+            tags: Additional tags to add to the run
+            tracking_uri: Override default tracking URI
+        """
+        self.config = MLflowConfig()
+        self.experiment_name = experiment_name
+        self.run_name = run_name or f"{experiment_name}-{int(time.time())}"
+        self.pipeline_metadata = pipeline_metadata
+        self.user_tags = tags or {}
+        self.tracking_uri = tracking_uri
+        
+        self.client: Optional[MlflowClient] = None
+        self.run: Optional[mlflow.ActiveRun] = None
+        self.run_id: Optional[str] = None
+        self._start_time: Optional[float] = None
+        
+    def _get_all_tags(self) -> Dict[str, str]:
+        """Combine all tags for the run."""
+        tags = self.config.default_tags.copy()
+        
+        if self.pipeline_metadata:
+            tags.update(self.pipeline_metadata.as_tags())
+        
+        tags.update(self.user_tags)
+        return tags
+    
+    @contextmanager
+    def start_run(
+        self,
+        nested: bool = False,
+        parent_run_id: Optional[str] = None,
+    ):
+        """
+        Start an MLflow run as a context manager.
+        
+        Args:
+            nested: If True, create a nested run under the current active run
+            parent_run_id: Explicit parent run ID for nested runs
+            
+        Yields:
+            The MLflow run object
+        """
+        self.client = get_mlflow_client(
+            tracking_uri=self.tracking_uri,
+            configure_global=True
+        )
+        
+        # Ensure experiment exists
+        experiment_id = ensure_experiment(self.experiment_name)
+        
+        self._start_time = time.time()
+        
+        try:
+            # Start the run
+            self.run = mlflow.start_run(
+                experiment_id=experiment_id,
+                run_name=self.run_name,
+                nested=nested,
+                tags=self._get_all_tags(),
+            )
+            self.run_id = self.run.info.run_id
+            
+            logger.info(
+                f"Started MLflow run '{self.run_name}' "
+                f"(ID: {self.run_id}) in experiment '{self.experiment_name}'"
+            )
+            
+            yield self.run
+            
+        except Exception as e:
+            logger.error(f"MLflow run failed: {e}")
+            if self.run:
+                mlflow.set_tag("run.status", "failed")
+                mlflow.set_tag("run.error", str(e))
+            raise
+        finally:
+            # Log duration
+            if self._start_time:
+                duration = time.time() - self._start_time
+                try:
+                    mlflow.log_metric("run_duration_seconds", duration)
+                except Exception:
+                    pass
+            
+            # End the run
+            mlflow.end_run()
+            logger.info(f"Ended MLflow run '{self.run_name}'")
+    
+    def log_params(self, params: Dict[str, Any]) -> None:
+        """
+        Log parameters to the current run.
+        
+        Args:
+            params: Dictionary of parameter names to values
+        """
+        if not self.run:
+            logger.warning("No active run, skipping log_params")
+            return
+        
+        # MLflow has limits on param values, truncate if needed
+        cleaned_params = {}
+        for key, value in params.items():
+            str_value = str(value)
+            if len(str_value) > 500:
+                str_value = str_value[:497] + "..."
+            cleaned_params[key] = str_value
+        
+        mlflow.log_params(cleaned_params)
+        logger.debug(f"Logged {len(params)} parameters")
+    
+    def log_param(self, key: str, value: Any) -> None:
+        """Log a single parameter."""
+        self.log_params({key: value})
+    
+    def log_metrics(
+        self,
+        metrics: Dict[str, Union[float, int]],
+        step: Optional[int] = None
+    ) -> None:
+        """
+        Log metrics to the current run.
+        
+        Args:
+            metrics: Dictionary of metric names to values
+            step: Optional step number for time-series metrics
+        """
+        if not self.run:
+            logger.warning("No active run, skipping log_metrics")
+            return
+        
+        mlflow.log_metrics(metrics, step=step)
+        logger.debug(f"Logged {len(metrics)} metrics")
+    
+    def log_metric(
+        self,
+        key: str,
+        value: Union[float, int],
+        step: Optional[int] = None
+    ) -> None:
+        """Log a single metric."""
+        self.log_metrics({key: value}, step=step)
+    
+    def log_artifact(
+        self,
+        local_path: str,
+        artifact_path: Optional[str] = None
+    ) -> None:
+        """
+        Log an artifact file to the current run.
+        
+        Args:
+            local_path: Path to the local file to log
+            artifact_path: Optional destination path within the artifact store
+        """
+        if not self.run:
+            logger.warning("No active run, skipping log_artifact")
+            return
+        
+        mlflow.log_artifact(local_path, artifact_path)
+        logger.info(f"Logged artifact: {local_path}")
+    
+    def log_artifacts(
+        self,
+        local_dir: str,
+        artifact_path: Optional[str] = None
+    ) -> None:
+        """
+        Log all files in a directory as artifacts.
+        
+        Args:
+            local_dir: Path to the local directory
+            artifact_path: Optional destination path within the artifact store
+        """
+        if not self.run:
+            logger.warning("No active run, skipping log_artifacts")
+            return
+        
+        mlflow.log_artifacts(local_dir, artifact_path)
+        logger.info(f"Logged artifacts from: {local_dir}")
+    
+    def log_dict(
+        self,
+        data: Dict[str, Any],
+        filename: str,
+        artifact_path: Optional[str] = None
+    ) -> None:
+        """
+        Log a dictionary as a JSON artifact.
+        
+        Args:
+            data: Dictionary to log
+            filename: Name for the JSON file
+            artifact_path: Optional destination path
+        """
+        if not self.run:
+            logger.warning("No active run, skipping log_dict")
+            return
+        
+        # Ensure .json extension
+        if not filename.endswith(".json"):
+            filename += ".json"
+        
+        mlflow.log_dict(data, f"{artifact_path}/{filename}" if artifact_path else filename)
+        logger.debug(f"Logged dict as: {filename}")
+    
+    def log_model_info(
+        self,
+        model_type: str,
+        model_name: str,
+        model_path: Optional[str] = None,
+        framework: str = "pytorch",
+        extra_info: Optional[Dict[str, Any]] = None
+    ) -> None:
+        """
+        Log model information as parameters and tags.
+        
+        Args:
+            model_type: Type of model (e.g., "llm", "embedding", "stt")
+            model_name: Name/identifier of the model
+            model_path: Path to model weights
+            framework: ML framework used
+            extra_info: Additional model information
+        """
+        params = {
+            "model.type": model_type,
+            "model.name": model_name,
+            "model.framework": framework,
+        }
+        if model_path:
+            params["model.path"] = model_path
+        if extra_info:
+            for key, value in extra_info.items():
+                params[f"model.{key}"] = value
+        
+        self.log_params(params)
+        
+        # Also set as tags for easier filtering
+        mlflow.set_tag("model.type", model_type)
+        mlflow.set_tag("model.name", model_name)
+    
+    def log_dataset_info(
+        self,
+        name: str,
+        source: str,
+        size: Optional[int] = None,
+        extra_info: Optional[Dict[str, Any]] = None
+    ) -> None:
+        """
+        Log dataset information.
+        
+        Args:
+            name: Dataset name
+            source: Dataset source (URL, path, etc.)
+            size: Number of samples
+            extra_info: Additional dataset information
+        """
+        params = {
+            "dataset.name": name,
+            "dataset.source": source,
+        }
+        if size is not None:
+            params["dataset.size"] = size
+        if extra_info:
+            for key, value in extra_info.items():
+                params[f"dataset.{key}"] = value
+        
+        self.log_params(params)
+    
+    def set_tag(self, key: str, value: str) -> None:
+        """Set a single tag on the run."""
+        if self.run:
+            mlflow.set_tag(key, value)
+    
+    def set_tags(self, tags: Dict[str, str]) -> None:
+        """Set multiple tags on the run."""
+        if self.run:
+            mlflow.set_tags(tags)
+    
+    @property
+    def artifact_uri(self) -> Optional[str]:
+        """Get the artifact URI for the current run."""
+        if self.run:
+            return self.run.info.artifact_uri
+        return None
+    
+    @property
+    def experiment_id(self) -> Optional[str]:
+        """Get the experiment ID for the current run."""
+        if self.run:
+            return self.run.info.experiment_id
+        return None
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 0000000..1828f85
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1,19 @@
+# MLflow Utils Module Requirements
+# Core MLflow
+mlflow>=2.10.0
+
+# Database backends
+psycopg2-binary>=2.9.0  # PostgreSQL (CNPG)
+boto3>=1.34.0  # S3-compatible artifact storage (optional)
+
+# For async tracking
+aiohttp>=3.9.0
+
+# YAML generation for KServe manifests
+PyYAML>=6.0
+
+# Already in chat-handler/voice-assistant requirements:
+# httpx (for health checks)
+# Used but typically installed with mlflow:
+# numpy
+# pandas