Change automated training interval from 2 minutes to 1 day

Fix root cause: Parse JSON strings from Supabase before accessing as dictionaries

Fix syntax error: correct indentation for try-except block in run_training_cycle

Add comprehensive error handling and traceback logging to identify str.get() error source

Deep fix: Add comprehensive state validation before logging training cycles to prevent type errors

Add automated training scheduler - checks every 2 minutes and trains when 10+ feedback samples available

Fix: safe extraction of version_id to avoid attribute errors when version state is string

Fix all dictionary access errors in feedback pipeline with proper type checks

Add graceful handling for missing database tables in model versioning

Fix error handling in feedback pipeline for model versioning

Add model versioning and training history tracking system

Add feedback learning pipeline for continuous model improvement

Add severity-based color coding for bounding boxes with confidence-scaled intensity

Add confidence scores to bounding box detections

Implement complete classification with all detection rules - Point/Tiny/Wire/Loose Joint detection

Fix image download to preserve format and add classification debugging

Add open-clip-torch dependency for anomalib VLM models

Fix Dockerfile: Add --user flag, ca-certificates, and improve gdown with --fuzzy flag

Fix: Add python-dotenv dependency and fix matplotlib permissions with non-root user

Optimize Docker image - Use CPU-only PyTorch and minimal dependencies to reduce size

Deploy Anomaly Detection API

initial commit