fix: remove unused imports and apply ruff formatting

- Remove unused imports: json (llm.py), tempfile (stt.py), base64 (tts.py)
- Apply ruff format to all Python files

.gitea/workflows/ci.yml

@@ -0,0 +1,240 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+env:
+  NTFY_URL: http://ntfy.observability.svc.cluster.local:80
+  REGISTRY: gitea-http.gitea.svc.cluster.local:3000/daviestechlabs
+  REGISTRY_HOST: gitea-http.gitea.svc.cluster.local:3000
+  IMAGE_NAME: gradio-ui
+  KUSTOMIZE_NAMESPACE: ai-ml
+
+jobs:
+  lint:
+    name: Lint
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up uv
+        run: curl -LsSf https://astral.sh/uv/install.sh | sh && echo "$HOME/.local/bin" >> $GITHUB_PATH
+
+      - name: Run ruff check
+        run: uvx ruff check .
+
+      - name: Run ruff format check
+        run: uvx ruff format --check .
+
+  release:
+    name: Release
+    runs-on: ubuntu-latest
+    needs: [lint]
+    if: gitea.ref == 'refs/heads/main' && gitea.event_name == 'push'
+    outputs:
+      version: ${{ steps.version.outputs.version }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Determine version bump
+        id: version
+        run: |
+          LATEST=$(git describe --tags --abbrev=0 2>/dev/null || echo "v0.0.0")
+          VERSION=${LATEST#v}
+          IFS='.' read -r MAJOR MINOR PATCH <<< "$VERSION"
+
+          MSG="${{ gitea.event.head_commit.message }}"
+          if echo "$MSG" | grep -qiE "^major:|BREAKING CHANGE"; then
+            MAJOR=$((MAJOR + 1)); MINOR=0; PATCH=0
+            BUMP="major"
+          elif echo "$MSG" | grep -qiE "^(minor:|feat:)"; then
+            MINOR=$((MINOR + 1)); PATCH=0
+            BUMP="minor"
+          else
+            PATCH=$((PATCH + 1))
+            BUMP="patch"
+          fi
+
+          NEW_VERSION="v${MAJOR}.${MINOR}.${PATCH}"
+          echo "version=$NEW_VERSION" >> $GITHUB_OUTPUT
+          echo "bump=$BUMP" >> $GITHUB_OUTPUT
+          echo "Bumping $LATEST → $NEW_VERSION ($BUMP)"
+
+      - name: Create and push tag
+        run: |
+          git config user.name "gitea-actions[bot]"
+          git config user.email "actions@git.daviestechlabs.io"
+          git tag -a ${{ steps.version.outputs.version }} -m "Release ${{ steps.version.outputs.version }}"
+          git push origin ${{ steps.version.outputs.version }}
+
+  docker:
+    name: Docker Build & Push
+    runs-on: ubuntu-latest
+    needs: [lint, release]
+    if: gitea.ref == 'refs/heads/main' && gitea.event_name == 'push'
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+        with:
+          buildkitd-config-inline: |
+            [registry."gitea-http.gitea.svc.cluster.local:3000"]
+              http = true
+              insecure = true
+
+      - name: Login to Docker Hub
+        if: vars.DOCKERHUB_USERNAME != ''
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+
+      - name: Configure Docker for insecure registry
+        run: |
+          sudo mkdir -p /etc/docker
+          echo '{"insecure-registries": ["${{ env.REGISTRY_HOST }}"]}' | sudo tee /etc/docker/daemon.json
+          sudo systemctl restart docker || sudo service docker restart || true
+          sleep 2
+
+      - name: Login to Gitea Registry
+        run: |
+          AUTH=$(echo -n "${{ secrets.REGISTRY_USER }}:${{ secrets.REGISTRY_TOKEN }}" | base64 -w0)
+          mkdir -p ~/.docker
+          cat > ~/.docker/config.json << EOF
+          {
+            "auths": {
+              "${{ env.REGISTRY_HOST }}": {
+                "auth": "$AUTH"
+              }
+            }
+          }
+          EOF
+          echo "Auth configured for ${{ env.REGISTRY_HOST }}"
+
+      - name: Extract metadata
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}
+          tags: |
+            type=semver,pattern={{version}},value=${{ needs.release.outputs.version }}
+            type=semver,pattern={{major}}.{{minor}},value=${{ needs.release.outputs.version }}
+            type=raw,value=latest,enable={{is_default_branch}}
+
+      - name: Build and push
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          push: true
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
+
+  deploy:
+    name: Deploy to Kubernetes
+    runs-on: ubuntu-latest
+    needs: [docker, release]
+    if: gitea.ref == 'refs/heads/main' && gitea.event_name == 'push'
+    container:
+      image: catthehacker/ubuntu:act-latest
+      volumes:
+        - /secrets/kubeconfig:/secrets/kubeconfig
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Install kubectl
+        run: |
+          curl -LO "https://dl.k8s.io/release/$(curl -Ls https://dl.k8s.io/release/stable.txt)/bin/linux/amd64/kubectl"
+          chmod +x kubectl && sudo mv kubectl /usr/local/bin/
+
+      - name: Update image tag in manifests
+        env:
+          KUBECONFIG: /secrets/kubeconfig/config
+        run: |
+          VERSION="${{ needs.release.outputs.version }}"
+          VERSION="${VERSION#v}"
+          for DEPLOY in llm embeddings stt tts; do
+            sed -i "s|image: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:.*|image: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${VERSION}|" "${DEPLOY}.yaml"
+          done
+
+      - name: Apply kustomization
+        env:
+          KUBECONFIG: /secrets/kubeconfig/config
+        run: |
+          kubectl apply -k . --namespace ${{ env.KUSTOMIZE_NAMESPACE }}
+
+      - name: Rollout restart deployments
+        env:
+          KUBECONFIG: /secrets/kubeconfig/config
+        run: |
+          for DEPLOY in llm-ui embeddings-ui stt-ui tts-ui; do
+            kubectl rollout restart deployment/${DEPLOY} -n ${{ env.KUSTOMIZE_NAMESPACE }} 2>/dev/null || true
+          done
+
+      - name: Wait for rollout
+        env:
+          KUBECONFIG: /secrets/kubeconfig/config
+        run: |
+          for DEPLOY in llm-ui embeddings-ui stt-ui tts-ui; do
+            kubectl rollout status deployment/${DEPLOY} -n ${{ env.KUSTOMIZE_NAMESPACE }} --timeout=120s 2>/dev/null || true
+          done
+
+  notify:
+    name: Notify
+    runs-on: ubuntu-latest
+    needs: [lint, release, docker, deploy]
+    if: always()
+    steps:
+      - name: Notify on success
+        if: needs.lint.result == 'success' && needs.docker.result == 'success'
+        run: |
+          curl -s \
+            -H "Title: ✅ CI Passed: ${{ gitea.repository }}" \
+            -H "Priority: default" \
+            -H "Tags: white_check_mark,github" \
+            -H "Click: ${{ gitea.server_url }}/${{ gitea.repository }}/actions/runs/${{ gitea.run_id }}" \
+            -d "Branch: ${{ gitea.ref_name }}
+          Commit: ${{ gitea.event.head_commit.message || gitea.sha }}
+          Release: ${{ needs.release.result == 'success' && needs.release.outputs.version || 'skipped' }}
+          Docker: ${{ needs.docker.result }}
+          Deploy: ${{ needs.deploy.result }}" \
+            ${{ env.NTFY_URL }}/gitea-ci
+
+      - name: Notify on deploy success
+        if: needs.deploy.result == 'success'
+        run: |
+          curl -s \
+            -H "Title: 🚀 Deployed: ${{ gitea.repository }}" \
+            -H "Priority: default" \
+            -H "Tags: rocket,kubernetes" \
+            -H "Click: ${{ gitea.server_url }}/${{ gitea.repository }}/actions/runs/${{ gitea.run_id }}" \
+            -d "Version: ${{ needs.release.outputs.version }}
+          Namespace: ${{ env.KUSTOMIZE_NAMESPACE }}
+          Apps: llm-ui, embeddings-ui, stt-ui, tts-ui" \
+            ${{ env.NTFY_URL }}/gitea-ci
+
+      - name: Notify on failure
+        if: needs.lint.result == 'failure' || needs.docker.result == 'failure' || needs.deploy.result == 'failure'
+        run: |
+          curl -s \
+            -H "Title: ❌ CI Failed: ${{ gitea.repository }}" \
+            -H "Priority: high" \
+            -H "Tags: x,github" \
+            -H "Click: ${{ gitea.server_url }}/${{ gitea.repository }}/actions/runs/${{ gitea.run_id }}" \
+            -d "Branch: ${{ gitea.ref_name }}
+          Commit: ${{ gitea.event.head_commit.message || gitea.sha }}
+          Lint: ${{ needs.lint.result }}
+          Docker: ${{ needs.docker.result }}
+          Deploy: ${{ needs.deploy.result }}" \
+            ${{ env.NTFY_URL }}/gitea-ci

embeddings.py

@@ -9,6 +9,7 @@ Features:
 - MLflow metrics logging
 - Visual embedding dimension display
 """
+
 import os
 import time
 import logging
@@ -28,12 +29,77 @@ logger = logging.getLogger("embeddings-demo")
 EMBEDDINGS_URL = os.environ.get(
     "EMBEDDINGS_URL",
     # Default: Ray Serve Embeddings endpoint
-    "http://ai-inference-serve-svc.ai-ml.svc.cluster.local:8000/embeddings"
+    "http://ai-inference-serve-svc.ai-ml.svc.cluster.local:8000/embeddings",
 )
+# ─── MLflow experiment tracking ──────────────────────────────────────────
+try:
+    import mlflow
+    from mlflow.tracking import MlflowClient
+
     MLFLOW_TRACKING_URI = os.environ.get(
         "MLFLOW_TRACKING_URI",
-    "http://mlflow.mlflow.svc.cluster.local:80"
+        "http://mlflow.mlflow.svc.cluster.local:80",
     )
+    mlflow.set_tracking_uri(MLFLOW_TRACKING_URI)
+    _mlflow_client = MlflowClient()
+
+    _experiment = _mlflow_client.get_experiment_by_name("gradio-embeddings-tuning")
+    if _experiment is None:
+        _experiment_id = _mlflow_client.create_experiment(
+            "gradio-embeddings-tuning",
+            artifact_location="/mlflow/artifacts/gradio-embeddings-tuning",
+        )
+    else:
+        _experiment_id = _experiment.experiment_id
+
+    _mlflow_run = mlflow.start_run(
+        experiment_id=_experiment_id,
+        run_name=f"gradio-embeddings-{os.environ.get('HOSTNAME', 'local')}",
+        tags={"service": "gradio-embeddings", "endpoint": EMBEDDINGS_URL},
+    )
+    _mlflow_run_id = _mlflow_run.info.run_id
+    _mlflow_step = 0
+    MLFLOW_ENABLED = True
+    logger.info(
+        "MLflow tracking enabled: experiment=%s run=%s", _experiment_id, _mlflow_run_id
+    )
+except Exception as exc:
+    logger.warning("MLflow tracking disabled: %s", exc)
+    _mlflow_client = None
+    _mlflow_run_id = None
+    _mlflow_step = 0
+    MLFLOW_ENABLED = False
+
+
+def _log_embedding_metrics(
+    latency: float, batch_size: int, embedding_dims: int = 0
+) -> None:
+    """Log embedding inference metrics to MLflow (non-blocking best-effort)."""
+    global _mlflow_step
+    if not MLFLOW_ENABLED or _mlflow_client is None:
+        return
+    try:
+        _mlflow_step += 1
+        ts = int(time.time() * 1000)
+        _mlflow_client.log_batch(
+            _mlflow_run_id,
+            metrics=[
+                mlflow.entities.Metric("latency_s", latency, ts, _mlflow_step),
+                mlflow.entities.Metric("batch_size", batch_size, ts, _mlflow_step),
+                mlflow.entities.Metric(
+                    "embedding_dims", embedding_dims, ts, _mlflow_step
+                ),
+                mlflow.entities.Metric(
+                    "latency_per_text_ms",
+                    (latency * 1000 / batch_size) if batch_size > 0 else 0,
+                    ts,
+                    _mlflow_step,
+                ),
+            ],
+        )
+    except Exception:
+        logger.debug("MLflow log failed", exc_info=True)
+
 
 # HTTP client
 client = httpx.Client(timeout=60.0)
@@ -44,8 +110,7 @@ def get_embeddings(texts: list[str]) -> tuple[list[list[float]], float]:
     start_time = time.time()
 
     response = client.post(
-        f"{EMBEDDINGS_URL}/embeddings",
+        f"{EMBEDDINGS_URL}/embeddings", json={"input": texts, "model": "bge"}
-        json={"input": texts, "model": "bge"}
     )
     response.raise_for_status()
 
@@ -77,6 +142,9 @@ def generate_single_embedding(text: str) -> tuple[str, str, str]:
         embedding = embeddings[0]
         dims = len(embedding)
 
+        # Log to MLflow
+        _log_embedding_metrics(latency, batch_size=1, embedding_dims=dims)
+
         # Format output
         status = f"✅ Generated {dims}-dimensional embedding in {latency * 1000:.1f}ms"
 
@@ -119,6 +187,9 @@ def compare_texts(text1: str, text2: str) -> tuple[str, str]:
 
         similarity = cosine_similarity(embeddings[0], embeddings[1])
 
+        # Log to MLflow
+        _log_embedding_metrics(latency, batch_size=2, embedding_dims=len(embeddings[0]))
+
         # Determine similarity level
         if similarity > 0.9:
             level = "🟢 Very High"
@@ -167,6 +238,13 @@ def batch_embed(texts_input: str) -> tuple[str, str]:
     try:
         embeddings, latency = get_embeddings(texts)
 
+        # Log to MLflow
+        _log_embedding_metrics(
+            latency,
+            batch_size=len(embeddings),
+            embedding_dims=len(embeddings[0]) if embeddings else 0,
+        )
+
         status = f"✅ Generated {len(embeddings)} embeddings in {latency * 1000:.1f}ms"
         status += f" ({latency * 1000 / len(texts):.1f}ms per text)"
 
@@ -243,7 +321,7 @@ Generate embeddings, compare text similarity, and explore vector representations
                     single_input = gr.Textbox(
                         label="Input Text",
                         placeholder="Enter text to generate embeddings...",
-                        lines=3
+                        lines=3,
                     )
                     single_btn = gr.Button("Generate Embedding", variant="primary")
 
@@ -256,7 +334,7 @@ Generate embeddings, compare text similarity, and explore vector representations
             single_btn.click(
                 fn=generate_single_embedding,
                 inputs=single_input,
-                outputs=[single_status, single_preview, single_stats]
+                outputs=[single_status, single_preview, single_stats],
             )
 
         # Tab 2: Compare Texts
@@ -276,14 +354,17 @@ Generate embeddings, compare text similarity, and explore vector representations
             compare_btn.click(
                 fn=compare_texts,
                 inputs=[compare_text1, compare_text2],
-                outputs=[compare_result, compare_visual]
+                outputs=[compare_result, compare_visual],
             )
 
             # Example pairs
             gr.Examples(
                 examples=[
                     ["The cat sat on the mat.", "A feline was resting on the rug."],
-                    ["Machine learning is a subset of AI.", "Deep learning uses neural networks."],
+                    [
+                        "Machine learning is a subset of AI.",
+                        "Deep learning uses neural networks.",
+                    ],
                     ["I love pizza.", "The stock market crashed today."],
                 ],
                 inputs=[compare_text1, compare_text2],
@@ -291,21 +372,21 @@ Generate embeddings, compare text similarity, and explore vector representations
 
         # Tab 3: Batch Embeddings
         with gr.TabItem("📚 Batch Processing"):
-            gr.Markdown("Generate embeddings for multiple texts and see their similarity matrix.")
+            gr.Markdown(
+                "Generate embeddings for multiple texts and see their similarity matrix."
+            )
 
             batch_input = gr.Textbox(
                 label="Texts (one per line)",
                 placeholder="Enter multiple texts, one per line...",
-                lines=6
+                lines=6,
             )
             batch_btn = gr.Button("Process Batch", variant="primary")
             batch_status = gr.Textbox(label="Status", interactive=False)
             batch_result = gr.Markdown(label="Similarity Matrix")
 
             batch_btn.click(
-                fn=batch_embed,
+                fn=batch_embed, inputs=batch_input, outputs=[batch_status, batch_result]
-                inputs=batch_input,
-                outputs=[batch_status, batch_result]
             )
 
             gr.Examples(
@@ -320,8 +401,4 @@ Generate embeddings, compare text similarity, and explore vector representations
 
 
 if __name__ == "__main__":
-    demo.launch(
+    demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True)
-        server_name="0.0.0.0",
-        server_port=7860,
-        show_error=True
-    )

embeddings.yaml

@@ -20,7 +20,7 @@ spec:
     spec:
       containers:
         - name: gradio
-          image: ghcr.io/billy-davies-2/llm-apps:v2-202602120526
+          image: gitea-http.gitea.svc.cluster.local:3000/daviestechlabs/gradio-ui:latest
           imagePullPolicy: Always
           command: ["python", "embeddings.py"]
           ports:

llm.py

@@ -9,10 +9,10 @@ Features:
 - Token usage and latency metrics
 - Chat history management
 """
+
 import os
 import time
 import logging
-import json
 
 import gradio as gr
 import httpx
@@ -30,6 +30,92 @@ LLM_URL = os.environ.get(
     "http://ai-inference-serve-svc.ai-ml.svc.cluster.local:8000/llm",
 )
 
+# ─── MLflow experiment tracking ──────────────────────────────────────────
+try:
+    import mlflow
+    from mlflow.tracking import MlflowClient
+
+    MLFLOW_TRACKING_URI = os.environ.get(
+        "MLFLOW_TRACKING_URI",
+        "http://mlflow.mlflow.svc.cluster.local:80",
+    )
+    mlflow.set_tracking_uri(MLFLOW_TRACKING_URI)
+    _mlflow_client = MlflowClient()
+
+    # Ensure experiment exists
+    _experiment = _mlflow_client.get_experiment_by_name("gradio-llm-tuning")
+    if _experiment is None:
+        _experiment_id = _mlflow_client.create_experiment(
+            "gradio-llm-tuning",
+            artifact_location="/mlflow/artifacts/gradio-llm-tuning",
+        )
+    else:
+        _experiment_id = _experiment.experiment_id
+
+    # One persistent run per Gradio instance
+    _mlflow_run = mlflow.start_run(
+        experiment_id=_experiment_id,
+        run_name=f"gradio-llm-{os.environ.get('HOSTNAME', 'local')}",
+        tags={
+            "service": "gradio-llm",
+            "endpoint": LLM_URL,
+            "mlflow.runName": f"gradio-llm-{os.environ.get('HOSTNAME', 'local')}",
+        },
+    )
+    _mlflow_run_id = _mlflow_run.info.run_id
+    _mlflow_step = 0
+    MLFLOW_ENABLED = True
+    logger.info(
+        "MLflow tracking enabled: experiment=%s run=%s", _experiment_id, _mlflow_run_id
+    )
+except Exception as exc:
+    logger.warning("MLflow tracking disabled: %s", exc)
+    _mlflow_client = None
+    _mlflow_run_id = None
+    _mlflow_step = 0
+    MLFLOW_ENABLED = False
+
+
+def _log_llm_metrics(
+    latency: float,
+    prompt_tokens: int,
+    completion_tokens: int,
+    temperature: float,
+    max_tokens: int,
+    top_p: float,
+) -> None:
+    """Log inference metrics to MLflow (non-blocking best-effort)."""
+    global _mlflow_step
+    if not MLFLOW_ENABLED or _mlflow_client is None:
+        return
+    try:
+        _mlflow_step += 1
+        ts = int(time.time() * 1000)
+        total_tokens = prompt_tokens + completion_tokens
+        tps = completion_tokens / latency if latency > 0 else 0
+        _mlflow_client.log_batch(
+            _mlflow_run_id,
+            metrics=[
+                mlflow.entities.Metric("latency_s", latency, ts, _mlflow_step),
+                mlflow.entities.Metric(
+                    "prompt_tokens", prompt_tokens, ts, _mlflow_step
+                ),
+                mlflow.entities.Metric(
+                    "completion_tokens", completion_tokens, ts, _mlflow_step
+                ),
+                mlflow.entities.Metric("total_tokens", total_tokens, ts, _mlflow_step),
+                mlflow.entities.Metric("tokens_per_second", tps, ts, _mlflow_step),
+                mlflow.entities.Metric("temperature", temperature, ts, _mlflow_step),
+                mlflow.entities.Metric(
+                    "max_tokens_requested", max_tokens, ts, _mlflow_step
+                ),
+                mlflow.entities.Metric("top_p", top_p, ts, _mlflow_step),
+            ],
+        )
+    except Exception:
+        logger.debug("MLflow log failed", exc_info=True)
+
+
 DEFAULT_SYSTEM_PROMPT = (
     "You are a helpful AI assistant running on Davies Tech Labs homelab infrastructure. "
     "You are powered by Llama 3.1 70B served via vLLM on AMD Strix Halo (ROCm). "
@@ -38,7 +124,7 @@ DEFAULT_SYSTEM_PROMPT = (
 
 # Use async client for streaming
 async_client = httpx.AsyncClient(timeout=httpx.Timeout(300.0, connect=30.0))
-sync_client = httpx.Client(timeout=10.0)
+sync_client = httpx.Client(timeout=httpx.Timeout(60.0, connect=10.0))
 
 
 async def chat_stream(
@@ -90,6 +176,16 @@ async def chat_stream(
             usage.get("completion_tokens", 0),
         )
 
+        # Log to MLflow
+        _log_llm_metrics(
+            latency=latency,
+            prompt_tokens=usage.get("prompt_tokens", 0),
+            completion_tokens=usage.get("completion_tokens", 0),
+            temperature=temperature,
+            max_tokens=max_tokens,
+            top_p=top_p,
+        )
+
         # Yield text progressively for a nicer streaming feel
         chunk_size = 4
         words = text.split(" ")
@@ -112,6 +208,13 @@ async def chat_stream(
 def check_service_health() -> str:
     """Check if the LLM service is reachable."""
     try:
+        # Try a lightweight GET against the Ray Serve base first.
+        # This avoids burning GPU time on a full inference round-trip.
+        base_url = LLM_URL.rsplit("/", 1)[0]  # strip /llm path
+        response = sync_client.get(f"{base_url}/-/routes")
+        if response.status_code == 200:
+            return "🟢 LLM service is healthy"
+        # Fall back to a minimal inference probe
         response = sync_client.post(
             LLM_URL,
             json={
@@ -125,6 +228,8 @@ def check_service_health() -> str:
         return f"🟡 LLM responded with status {response.status_code}"
     except httpx.ConnectError:
         return "🔴 Cannot connect to LLM service"
+    except httpx.TimeoutException:
+        return "🟡 LLM service is reachable but slow to respond"
     except Exception as e:
         return f"🔴 Service unavailable: {e}"
 
@@ -164,13 +269,23 @@ def single_prompt(
         text = result["choices"][0]["message"]["content"]
         usage = result.get("usage", {})
 
+        # Log to MLflow
+        _log_llm_metrics(
+            latency=latency,
+            prompt_tokens=usage.get("prompt_tokens", 0),
+            completion_tokens=usage.get("completion_tokens", 0),
+            temperature=temperature,
+            max_tokens=max_tokens,
+            top_p=top_p,
+        )
+
         metrics = f"""
 **Generation Metrics:**
 - Latency: {latency:.1f}s
-- Prompt tokens: {usage.get('prompt_tokens', 'N/A')}
+- Prompt tokens: {usage.get("prompt_tokens", "N/A")}
-- Completion tokens: {usage.get('completion_tokens', 'N/A')}
+- Completion tokens: {usage.get("completion_tokens", "N/A")}
-- Total tokens: {usage.get('total_tokens', 'N/A')}
+- Total tokens: {usage.get("total_tokens", "N/A")}
-- Model: {result.get('model', 'N/A')}
+- Model: {result.get("model", "N/A")}
 """
         return text, metrics
 
@@ -218,18 +333,15 @@ Chat with **Llama 3.1 70B** (AWQ INT4) served via vLLM on AMD Strix Halo (ROCm).
         with gr.TabItem("💬 Chat"):
             chatbot = gr.ChatInterface(
                 fn=chat_stream,
-                type="messages",
                 additional_inputs=[system_prompt, temperature, max_tokens, top_p],
                 examples=[
-                    "Hello! What can you tell me about yourself?",
+                    ["Hello! What can you tell me about yourself?"],
-                    "Explain how a GPU executes a matrix multiplication.",
+                    ["Explain how a GPU executes a matrix multiplication."],
-                    "Write a Python function to compute the Fibonacci sequence.",
+                    ["Write a Python function to compute the Fibonacci sequence."],
-                    "What are the pros and cons of running LLMs on AMD GPUs?",
+                    ["What are the pros and cons of running LLMs on AMD GPUs?"],
                 ],
                 chatbot=gr.Chatbot(
                     height=520,
-                    type="messages",
-                    show_copy_button=True,
                     placeholder="Type a message to start chatting...",
                 ),
             )
@@ -257,9 +369,13 @@ Chat with **Llama 3.1 70B** (AWQ INT4) served via vLLM on AMD Strix Halo (ROCm).
 
             gr.Examples(
                 examples=[
-                    ["Summarise the key differences between CUDA and ROCm for ML workloads."],
+                    [
+                        "Summarise the key differences between CUDA and ROCm for ML workloads."
+                    ],
                     ["Write a haiku about Kubernetes."],
-                    ["Explain Ray Serve in one paragraph for someone new to ML serving."],
+                    [
+                        "Explain Ray Serve in one paragraph for someone new to ML serving."
+                    ],
                     ["List 5 creative uses for a homelab GPU cluster."],
                 ],
                 inputs=[prompt_input],

llm.yaml

@@ -20,7 +20,7 @@ spec:
     spec:
       containers:
         - name: gradio
-          image: ghcr.io/billy-davies-2/llm-apps:v2-202602120526
+          image: gitea-http.gitea.svc.cluster.local:3000/daviestechlabs/gradio-ui:latest
           imagePullPolicy: Always
           command: ["python", "llm.py"]
           ports:
@@ -53,7 +53,7 @@ spec:
             initialDelaySeconds: 5
             periodSeconds: 10
       imagePullSecrets:
-        - name: ghcr-registry
+        - name: gitea-registry
 ---
 apiVersion: v1
 kind: Service

renovate.json

@@ -0,0 +1,7 @@
+{
+  "$schema": "https://docs.renovatebot.com/renovate-schema.json",
+  "extends": [
+    "local>daviestechlabs/renovate-config",
+    "local>daviestechlabs/renovate-config:python"
+  ]
+}

stt.py

@@ -9,11 +9,11 @@ Features:
 - Translation mode
 - MLflow metrics logging
 """
+
 import os
 import time
 import logging
 import io
-import tempfile
 
 import gradio as gr
 import httpx
@@ -30,13 +30,82 @@ logger = logging.getLogger("stt-demo")
 STT_URL = os.environ.get(
     "STT_URL",
     # Default: Ray Serve whisper endpoint
-    "http://ai-inference-serve-svc.ai-ml.svc.cluster.local:8000/whisper"
+    "http://ai-inference-serve-svc.ai-ml.svc.cluster.local:8000/whisper",
 )
 MLFLOW_TRACKING_URI = os.environ.get(
-    "MLFLOW_TRACKING_URI",
+    "MLFLOW_TRACKING_URI", "http://mlflow.mlflow.svc.cluster.local:80"
-    "http://mlflow.mlflow.svc.cluster.local:80"
 )
 
+# ─── MLflow experiment tracking ──────────────────────────────────────────
+try:
+    import mlflow
+    from mlflow.tracking import MlflowClient
+
+    mlflow.set_tracking_uri(MLFLOW_TRACKING_URI)
+    _mlflow_client = MlflowClient()
+
+    _experiment = _mlflow_client.get_experiment_by_name("gradio-stt-tuning")
+    if _experiment is None:
+        _experiment_id = _mlflow_client.create_experiment(
+            "gradio-stt-tuning",
+            artifact_location="/mlflow/artifacts/gradio-stt-tuning",
+        )
+    else:
+        _experiment_id = _experiment.experiment_id
+
+    _mlflow_run = mlflow.start_run(
+        experiment_id=_experiment_id,
+        run_name=f"gradio-stt-{os.environ.get('HOSTNAME', 'local')}",
+        tags={"service": "gradio-stt", "endpoint": STT_URL},
+    )
+    _mlflow_run_id = _mlflow_run.info.run_id
+    _mlflow_step = 0
+    MLFLOW_ENABLED = True
+    logger.info(
+        "MLflow tracking enabled: experiment=%s run=%s", _experiment_id, _mlflow_run_id
+    )
+except Exception as exc:
+    logger.warning("MLflow tracking disabled: %s", exc)
+    _mlflow_client = None
+    _mlflow_run_id = None
+    _mlflow_step = 0
+    MLFLOW_ENABLED = False
+
+
+def _log_stt_metrics(
+    latency: float,
+    audio_duration: float,
+    word_count: int,
+    task: str,
+) -> None:
+    """Log STT inference metrics to MLflow (non-blocking best-effort)."""
+    global _mlflow_step
+    if not MLFLOW_ENABLED or _mlflow_client is None:
+        return
+    try:
+        _mlflow_step += 1
+        ts = int(time.time() * 1000)
+        rtf = latency / audio_duration if audio_duration > 0 else 0
+        _mlflow_client.log_batch(
+            _mlflow_run_id,
+            metrics=[
+                mlflow.entities.Metric("latency_s", latency, ts, _mlflow_step),
+                mlflow.entities.Metric(
+                    "audio_duration_s", audio_duration, ts, _mlflow_step
+                ),
+                mlflow.entities.Metric("realtime_factor", rtf, ts, _mlflow_step),
+                mlflow.entities.Metric("word_count", word_count, ts, _mlflow_step),
+            ],
+            params=[]
+            if _mlflow_step > 1
+            else [
+                mlflow.entities.Param("task", task),
+            ],
+        )
+    except Exception:
+        logger.debug("MLflow log failed", exc_info=True)
+
+
 # HTTP client with longer timeout for transcription
 client = httpx.Client(timeout=180.0)
 
@@ -63,9 +132,7 @@ LANGUAGES = {
 
 
 def transcribe_audio(
-    audio_input: tuple[int, np.ndarray] | str | None,
+    audio_input: tuple[int, np.ndarray] | str | None, language: str, task: str
-    language: str,
-    task: str
 ) -> tuple[str, str, str]:
     """Transcribe audio using the Whisper STT service."""
     if audio_input is None:
@@ -81,12 +148,12 @@ def transcribe_audio(
 
             # Convert to WAV bytes
             audio_buffer = io.BytesIO()
-            sf.write(audio_buffer, audio_data, sample_rate, format='WAV')
+            sf.write(audio_buffer, audio_data, sample_rate, format="WAV")
             audio_bytes = audio_buffer.getvalue()
             audio_duration = len(audio_data) / sample_rate
         else:
             # File path
-            with open(audio_input, 'rb') as f:
+            with open(audio_input, "rb") as f:
                 audio_bytes = f.read()
             # Get duration
             audio_data, sample_rate = sf.read(audio_input)
@@ -117,8 +184,18 @@ def transcribe_audio(
         text = result.get("text", "")
         detected_language = result.get("language", "unknown")
 
+        # Log to MLflow
+        _log_stt_metrics(
+            latency=latency,
+            audio_duration=audio_duration,
+            word_count=len(text.split()),
+            task=task,
+        )
+
         # Status message
-        status = f"✅ Transcribed {audio_duration:.1f}s of audio in {latency*1000:.0f}ms"
+        status = (
+            f"✅ Transcribed {audio_duration:.1f}s of audio in {latency * 1000:.0f}ms"
+        )
 
         # Metrics
         metrics = f"""
@@ -181,21 +258,19 @@ or file upload with support for 100+ languages.
             with gr.Row():
                 with gr.Column():
                     mic_input = gr.Audio(
-                        label="Record Audio",
+                        label="Record Audio", sources=["microphone"], type="numpy"
-                        sources=["microphone"],
-                        type="numpy"
                     )
 
                     with gr.Row():
                         mic_language = gr.Dropdown(
                             choices=list(LANGUAGES.keys()),
                             value="Auto-detect",
-                            label="Language"
+                            label="Language",
                         )
                         mic_task = gr.Radio(
                             choices=["Transcribe", "Translate to English"],
                             value="Transcribe",
-                            label="Task"
+                            label="Task",
                         )
 
                     mic_btn = gr.Button("🎯 Transcribe", variant="primary")
@@ -204,15 +279,12 @@ or file upload with support for 100+ languages.
                     mic_status = gr.Textbox(label="Status", interactive=False)
                     mic_metrics = gr.Markdown(label="Metrics")
 
-            mic_output = gr.Textbox(
+            mic_output = gr.Textbox(label="Transcription", lines=5)
-                label="Transcription",
-                lines=5
-            )
 
             mic_btn.click(
                 fn=transcribe_audio,
                 inputs=[mic_input, mic_language, mic_task],
-                outputs=[mic_status, mic_output, mic_metrics]
+                outputs=[mic_status, mic_output, mic_metrics],
             )
 
         # Tab 2: File Upload
@@ -220,21 +292,19 @@ or file upload with support for 100+ languages.
             with gr.Row():
                 with gr.Column():
                     file_input = gr.Audio(
-                        label="Upload Audio File",
+                        label="Upload Audio File", sources=["upload"], type="filepath"
-                        sources=["upload"],
-                        type="filepath"
                     )
 
                     with gr.Row():
                         file_language = gr.Dropdown(
                             choices=list(LANGUAGES.keys()),
                             value="Auto-detect",
-                            label="Language"
+                            label="Language",
                         )
                         file_task = gr.Radio(
                             choices=["Transcribe", "Translate to English"],
                             value="Transcribe",
-                            label="Task"
+                            label="Task",
                         )
 
                     file_btn = gr.Button("🎯 Transcribe", variant="primary")
@@ -243,15 +313,12 @@ or file upload with support for 100+ languages.
                     file_status = gr.Textbox(label="Status", interactive=False)
                     file_metrics = gr.Markdown(label="Metrics")
 
-            file_output = gr.Textbox(
+            file_output = gr.Textbox(label="Transcription", lines=5)
-                label="Transcription",
-                lines=5
-            )
 
             file_btn.click(
                 fn=transcribe_audio,
                 inputs=[file_input, file_language, file_task],
-                outputs=[file_status, file_output, file_metrics]
+                outputs=[file_status, file_output, file_metrics],
             )
 
             gr.Markdown("""
@@ -274,7 +341,7 @@ Whisper will automatically detect the source language.
                     trans_input = gr.Audio(
                         label="Audio Input",
                         sources=["microphone", "upload"],
-                        type="numpy"
+                        type="numpy",
                     )
                     trans_btn = gr.Button("🌍 Translate to English", variant="primary")
 
@@ -282,10 +349,7 @@ Whisper will automatically detect the source language.
                     trans_status = gr.Textbox(label="Status", interactive=False)
                     trans_metrics = gr.Markdown(label="Metrics")
 
-            trans_output = gr.Textbox(
+            trans_output = gr.Textbox(label="English Translation", lines=5)
-                label="English Translation",
-                lines=5
-            )
 
             def translate_audio(audio):
                 return transcribe_audio(audio, "Auto-detect", "Translate to English")
@@ -293,15 +357,11 @@ Whisper will automatically detect the source language.
             trans_btn.click(
                 fn=translate_audio,
                 inputs=trans_input,
-                outputs=[trans_status, trans_output, trans_metrics]
+                outputs=[trans_status, trans_output, trans_metrics],
             )
 
     create_footer()
 
 
 if __name__ == "__main__":
-    demo.launch(
+    demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True)
-        server_name="0.0.0.0",
-        server_port=7860,
-        show_error=True
-    )

stt.yaml

@@ -20,7 +20,7 @@ spec:
     spec:
       containers:
         - name: gradio
-          image: ghcr.io/billy-davies-2/llm-apps:v2-202602120526
+          image: gitea-http.gitea.svc.cluster.local:3000/daviestechlabs/gradio-ui:latest
           imagePullPolicy: Always
           command: ["python", "stt.py"]
           ports:

theme.py

@@ -3,6 +3,7 @@ Shared Gradio theme for Davies Tech Labs AI demos.
 Consistent styling across all demo applications.
 Cyberpunk aesthetic - dark with yellow/gold accents.
 """
+
 import gradio as gr
 
 
@@ -25,7 +26,12 @@ def get_lab_theme() -> gr.Theme:
         primary_hue=gr.themes.colors.yellow,
         secondary_hue=gr.themes.colors.amber,
         neutral_hue=gr.themes.colors.zinc,
-        font=[gr.themes.GoogleFont("Space Grotesk"), "ui-sans-serif", "system-ui", "sans-serif"],
+        font=[
+            gr.themes.GoogleFont("Space Grotesk"),
+            "ui-sans-serif",
+            "system-ui",
+            "sans-serif",
+        ],
         font_mono=[gr.themes.GoogleFont("JetBrains Mono"), "ui-monospace", "monospace"],
     ).set(
         # Background colors

tts.py

@@ -9,11 +9,11 @@ Features:
 - MLflow metrics logging
 - Multiple TTS backends support (Coqui XTTS, Piper, etc.)
 """
+
 import os
 import time
 import logging
 import io
-import base64
 
 import gradio as gr
 import httpx
@@ -30,13 +30,79 @@ logger = logging.getLogger("tts-demo")
 TTS_URL = os.environ.get(
     "TTS_URL",
     # Default: Ray Serve TTS endpoint
-    "http://ai-inference-serve-svc.ai-ml.svc.cluster.local:8000/tts"
+    "http://ai-inference-serve-svc.ai-ml.svc.cluster.local:8000/tts",
 )
 MLFLOW_TRACKING_URI = os.environ.get(
-    "MLFLOW_TRACKING_URI",
+    "MLFLOW_TRACKING_URI", "http://mlflow.mlflow.svc.cluster.local:80"
-    "http://mlflow.mlflow.svc.cluster.local:80"
 )
 
+# ─── MLflow experiment tracking ──────────────────────────────────────────
+try:
+    import mlflow
+    from mlflow.tracking import MlflowClient
+
+    mlflow.set_tracking_uri(MLFLOW_TRACKING_URI)
+    _mlflow_client = MlflowClient()
+
+    _experiment = _mlflow_client.get_experiment_by_name("gradio-tts-tuning")
+    if _experiment is None:
+        _experiment_id = _mlflow_client.create_experiment(
+            "gradio-tts-tuning",
+            artifact_location="/mlflow/artifacts/gradio-tts-tuning",
+        )
+    else:
+        _experiment_id = _experiment.experiment_id
+
+    _mlflow_run = mlflow.start_run(
+        experiment_id=_experiment_id,
+        run_name=f"gradio-tts-{os.environ.get('HOSTNAME', 'local')}",
+        tags={"service": "gradio-tts", "endpoint": TTS_URL},
+    )
+    _mlflow_run_id = _mlflow_run.info.run_id
+    _mlflow_step = 0
+    MLFLOW_ENABLED = True
+    logger.info(
+        "MLflow tracking enabled: experiment=%s run=%s", _experiment_id, _mlflow_run_id
+    )
+except Exception as exc:
+    logger.warning("MLflow tracking disabled: %s", exc)
+    _mlflow_client = None
+    _mlflow_run_id = None
+    _mlflow_step = 0
+    MLFLOW_ENABLED = False
+
+
+def _log_tts_metrics(
+    latency: float,
+    audio_duration: float,
+    text_chars: int,
+    language: str,
+) -> None:
+    """Log TTS inference metrics to MLflow (non-blocking best-effort)."""
+    global _mlflow_step
+    if not MLFLOW_ENABLED or _mlflow_client is None:
+        return
+    try:
+        _mlflow_step += 1
+        ts = int(time.time() * 1000)
+        rtf = latency / audio_duration if audio_duration > 0 else 0
+        cps = text_chars / latency if latency > 0 else 0
+        _mlflow_client.log_batch(
+            _mlflow_run_id,
+            metrics=[
+                mlflow.entities.Metric("latency_s", latency, ts, _mlflow_step),
+                mlflow.entities.Metric(
+                    "audio_duration_s", audio_duration, ts, _mlflow_step
+                ),
+                mlflow.entities.Metric("realtime_factor", rtf, ts, _mlflow_step),
+                mlflow.entities.Metric("chars_per_second", cps, ts, _mlflow_step),
+                mlflow.entities.Metric("text_chars", text_chars, ts, _mlflow_step),
+            ],
+        )
+    except Exception:
+        logger.debug("MLflow log failed", exc_info=True)
+
+
 # HTTP client with longer timeout for audio generation
 client = httpx.Client(timeout=120.0)
 
@@ -61,7 +127,9 @@ LANGUAGES = {
 }
 
 
-def synthesize_speech(text: str, language: str) -> tuple[str, tuple[int, np.ndarray] | None, str]:
+def synthesize_speech(
+    text: str, language: str
+) -> tuple[str, tuple[int, np.ndarray] | None, str]:
     """Synthesize speech from text using the TTS service."""
     if not text.strip():
         return "❌ Please enter some text", None, ""
@@ -73,8 +141,7 @@ def synthesize_speech(text: str, language: str) -> tuple[str, tuple[int, np.ndar
 
         # Call TTS service (Coqui XTTS API format)
         response = client.get(
-            f"{TTS_URL}/api/tts",
+            f"{TTS_URL}/api/tts", params={"text": text, "language_id": lang_code}
-            params={"text": text, "language_id": lang_code}
         )
         response.raise_for_status()
 
@@ -94,6 +161,14 @@ def synthesize_speech(text: str, language: str) -> tuple[str, tuple[int, np.ndar
         # Status message
         status = f"✅ Generated {duration:.2f}s of audio in {latency * 1000:.0f}ms"
 
+        # Log to MLflow
+        _log_tts_metrics(
+            latency=latency,
+            audio_duration=duration,
+            text_chars=len(text),
+            language=lang_code,
+        )
+
         # Metrics
         metrics = f"""
 **Audio Statistics:**
@@ -160,16 +235,18 @@ in multiple languages.
                         label="Text to Synthesize",
                         placeholder="Enter text to convert to speech...",
                         lines=5,
-                        max_lines=10
+                        max_lines=10,
                     )
 
                     with gr.Row():
                         language = gr.Dropdown(
                             choices=list(LANGUAGES.keys()),
                             value="English",
-                            label="Language"
+                            label="Language",
+                        )
+                        synthesize_btn = gr.Button(
+                            "🔊 Synthesize", variant="primary", scale=2
                         )
-                        synthesize_btn = gr.Button("🔊 Synthesize", variant="primary", scale=2)
 
                 with gr.Column(scale=1):
                     status_output = gr.Textbox(label="Status", interactive=False)
@@ -180,15 +257,24 @@ in multiple languages.
             synthesize_btn.click(
                 fn=synthesize_speech,
                 inputs=[text_input, language],
-                outputs=[status_output, audio_output, metrics_output]
+                outputs=[status_output, audio_output, metrics_output],
             )
 
             # Example texts
             gr.Examples(
                 examples=[
-                    ["Hello! Welcome to Davies Tech Labs. This is a demonstration of our text-to-speech system.", "English"],
+                    [
-                    ["The quick brown fox jumps over the lazy dog. This sentence contains every letter of the alphabet.", "English"],
+                        "Hello! Welcome to Davies Tech Labs. This is a demonstration of our text-to-speech system.",
-                    ["Bonjour! Bienvenue au laboratoire technique de Davies.", "French"],
+                        "English",
+                    ],
+                    [
+                        "The quick brown fox jumps over the lazy dog. This sentence contains every letter of the alphabet.",
+                        "English",
+                    ],
+                    [
+                        "Bonjour! Bienvenue au laboratoire technique de Davies.",
+                        "French",
+                    ],
                     ["Hola! Bienvenido al laboratorio de tecnología.", "Spanish"],
                     ["Guten Tag! Willkommen im Techniklabor.", "German"],
                 ],
@@ -200,14 +286,16 @@ in multiple languages.
             gr.Markdown("Compare the same text in different languages.")
 
             compare_text = gr.Textbox(
-                label="Text to Compare",
+                label="Text to Compare", value="Hello, how are you today?", lines=2
-                value="Hello, how are you today?",
-                lines=2
             )
 
             with gr.Row():
-                lang1 = gr.Dropdown(choices=list(LANGUAGES.keys()), value="English", label="Language 1")
+                lang1 = gr.Dropdown(
-                lang2 = gr.Dropdown(choices=list(LANGUAGES.keys()), value="Spanish", label="Language 2")
+                    choices=list(LANGUAGES.keys()), value="English", label="Language 1"
+                )
+                lang2 = gr.Dropdown(
+                    choices=list(LANGUAGES.keys()), value="Spanish", label="Language 2"
+                )
 
             compare_btn = gr.Button("Compare Languages", variant="primary")
 
@@ -230,7 +318,7 @@ in multiple languages.
             compare_btn.click(
                 fn=compare_languages,
                 inputs=[compare_text, lang1, lang2],
-                outputs=[status1, audio1, status2, audio2]
+                outputs=[status1, audio1, status2, audio2],
             )
 
         # Tab 3: Batch Processing
@@ -240,19 +328,16 @@ in multiple languages.
             batch_input = gr.Textbox(
                 label="Texts (one per line)",
                 placeholder="Enter multiple texts, one per line...",
-                lines=6
+                lines=6,
             )
             batch_lang = gr.Dropdown(
-                choices=list(LANGUAGES.keys()),
+                choices=list(LANGUAGES.keys()), value="English", label="Language"
-                value="English",
-                label="Language"
             )
             batch_btn = gr.Button("Synthesize All", variant="primary")
 
             batch_status = gr.Textbox(label="Status", interactive=False)
             batch_audios = gr.Dataset(
-                components=[gr.Audio(type="numpy")],
+                components=[gr.Audio(type="numpy")], label="Generated Audio Files"
-                label="Generated Audio Files"
             )
 
             # Note: Batch processing would need more complex handling
@@ -266,8 +351,4 @@ or the Kubeflow pipeline for better throughput.*
 
 
 if __name__ == "__main__":
-    demo.launch(
+    demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True)
-        server_name="0.0.0.0",
-        server_port=7860,
-        show_error=True
-    )

tts.yaml

@@ -20,7 +20,7 @@ spec:
     spec:
       containers:
         - name: gradio
-          image: ghcr.io/billy-davies-2/llm-apps:v2-202602120526
+          image: gitea-http.gitea.svc.cluster.local:3000/daviestechlabs/gradio-ui:latest
           imagePullPolicy: Always
           command: ["python", "tts.py"]
           ports: