Deployed 5-model computer vision pipeline (Mask R-CNN, 2 versions of YOLOv8, GradientBoosting, Gemini 2.5 Flash) achieving near production-grade metrics for automated vehicle damage assessment, cost estimation by region, and straight-through claims processing — containerized with Docker and deployed to production on Hugging Face Spaces and Vercel.
The frontend renders the uploaded image with color-coded bounding box overlays (yellow = minor, orange = moderate, red = severe) alongside part-level detections with confidence scores, severity ratings, per-part cost ranges, a natural language explanation, and fraud signal flags. After analysis, adjusters can change the state to instantly re-price all parts using that region's labor rates, or manually override individual part detections and costs. Session-based claim history is stored in an append-only audit log and accessible via a slide-in sidebar with CSV export.
State-based labor rates — select a state before upload to apply SCRS 2024 regional rates; body rates range from ~$59/hr (Southeast) to ~$84/hr (West Coast)
Live cost re-estimation — change the state dropdown in the results panel to instantly re-price all detected parts without re-uploading
Adjuster overrides — edit any part's detection (part, damage type, severity) and get a backend-recalculated cost range; override takes priority over state adjustments
Fraud signals — three passive checks (pixel variance, EXIF editing software, duplicate hash) flagged on every submission
Session claim history — each browser session has a unique ID; all claims for the session are viewable in a slide-in sidebar and exportable as CSV
Append-only audit log — every claim is logged with full model inputs/outputs for compliance review
