Fix: Case-insensitive expression key lookup for accurate percentages

Fix: Correct facial expression data extraction from service

Fix: Update AnalysisResult model to match new response format

Fix: Remove total_frames from video_info - not in VideoMetadata model

Fix: KeyError on expression dict access in first impression

Fix: Calculate total_frames from duration * fps

Implement Clean Response Structure with First Impression Analysis

Optimize Eye Tracking: Fix over-sensitivity + accurate gaze away calculation

Add flexible GPU/CPU runtime for MediaPipe (Eye Tracking + Gesture Detection)

Fix: Add onnxruntime-gpu + flexible ONNX execution providers

Fix: Disable ONNX Runtime auto-install, force PyTorch backend

Fix: ONNX model GPU inference - use device parameter instead of .to()

Enable GPU inference for facial expression on Nvidia T4

Fix: ONNX Runtime missing for facial expression on GPU Space

Fix: Permission denied error for logs directory on GPU Space

Remove frame skipping and debug video output to save storage

Fix: Complete NumPy bool serialization + result null issue

Fix: Adaptive frame sampling for long videos + NumPy bool serialization

Fix: Use fresh MediaPipe Pose instance per video to avoid timeout state

Fix: Add MediaPipe error handling + better error protection

Fix: Add frame sampling to gesture detection + simplify opening expressions to percentage

Fix: Enable eye tracking for Level 1 + optimize facial expression with frame sampling

Fix: Sequential processing to prevent OOM + increase timeout to 15min

Fix: Use decode_responses=False for RQ 2.0 worker compatibility

Fix: Remove deprecated Connection import from rq 2.0

Fix: Add Redis timeout config + subprocess text mode

Fix: Stream worker logs to stdout for visibility

Fix worker startup: Use Python script instead of bash to ensure worker runs

Fix: Run worker alongside API server to process background jobs

Trigger rebuild

Initial deployment: SWARA API with eye tracking, facial expression, and gesture detection

initial commit