Integrate ML Pipeline with Main Application

[machug]

Overview

The ML pipeline is complete and tested (PR #61) but not yet integrated with the main application. This issue tracks connecting the ML components to the live reading pipeline.

Current State

✅ Complete:

• ML pipeline implementation in backend/ml/
• 48/48 tests passing
• Per-device state isolation
• Kalman filtering, anomaly detection, curve fitting, MPC

❌ Not Connected:

• Not called from backend/main.py reading handler
• Not exposed via API endpoints
• Not displayed in frontend UI
• State lost on service restart

Proposed Integration

Phase 1: Basic Integration

1. Add MLPipelineManager to backend/main.py
2. Process readings through ML pipeline in handle_tilt_reading()
3. Broadcast ML results via WebSocket
4. Display filtered values on dashboard

Phase 2: State Persistence

5. Hydrate ML state from database on startup (last 48 hours)
6. Ensures predictions work immediately after restart
7. Option: Store ML results in new ml_results table

Phase 3: UI Integration

8. Show Kalman-filtered values on dashboard
9. Display FG predictions and completion estimates
10. Anomaly alerts (stuck fermentation, temperature spikes)
11. MPC temperature trajectory visualization

Data Storage Strategy

Current approach: In-Memory

• ML state lives in RAM (per-device history buffers)
• ~157 KB per device for 2-week fermentation
• Lost on service restart

Recommended: Hydrate from DB

Implementation Notes

• Keep ML processing async/non-blocking
• Feature flags for gradual rollout (MLConfig)
• WebSocket message type: ml_update
• Consider caching filtered values for API endpoints

Related

• PR feat: ML pipeline for fermentation analytics #61 - ML Pipeline Implementation
• Validation guide: ML_VALIDATION_GUIDE.md

Acceptance Criteria

• ML pipeline processes all live Tilt readings
• Filtered values broadcast via WebSocket
• ML state survives service restarts
• Dashboard shows filtered SG/temp
• Predictions visible on batch detail page
• Anomaly alerts displayed to user
• No performance degradation (<10ms overhead)

Priority

Medium - ML pipeline is functional but not user-visible. Can be done incrementally after PR #61 merges.