Spaces:
Sleeping
Sleeping
Commit History
-
fix: denomination-unconfirmed genuine verdicts now disclose withdrawn/demonetized risk instead of going silent
Hikitye41fdf9 -
fix: OCR rupee-symbol-misread-as-8 denomination fix + frozen ₹2000 regression test
Hikityc5a4ecc
-
fix: serial check demoted to corroboration-required + tight panel crop — genuine-note false positive eliminated
Hikity0a96cf6 -
feat: implement currency detection logic and document training datasets
Hikity963607c -
feat: add v2.2 currency detection pipeline and update health endpoint to support 8 total endpoints
Hikityadbbe6a -
feat: denomination-validity structural check + corrected real eval metrics (91.9% overall, 93.0% circulating avg)
Hikity673a9be -
docs: circulating-denomination framing — ₹10-₹500 average 99.7%, ₹2000 withdrawn
Hikity40c7db6 -
docs: counterfeit detection pipeline doc (flow diagram, PS mapping, model training story)
Hikitye985ce9 -
fix: warm EasyOCR at startup — cold-init on first currency request exceeded HF proxy timeout and 500'd after every container restart
Hikity717e080 -
feat: implement multi-stage currency counterfeit detection pipeline with OCR and structural analysis
Hikity2543a43 -
feat: implement KAWACH AI Classifier v2.1 service with six multi-modal pipelines and FastAPI integration
Hikitye2c7aa0
-
feat: add Pipeline 7 for counterfeit currency detection and upgrade to Gemini 2.5-flash
Hikity8adbf12 -
feat: implement currency classification pipeline with EfficientNet-B0 and model training utilities
Hikity2ba1249
-
feat: implement core FastAPI application structure, routing logic, and Pydantic schemas for the AI classification service.
Hikityb08ad52
-
docs: complete video processing flow diagram + component breakdown in README
Hikity +1: Claude Sonnet 4.6• Full ASCII flow diagram showing exactly what each Python file does at each stage: video_reader → face_extractor → classifier → verdict router → priority_validator → routing result scene_analyzer → YOLO/TrashNet per-frame → temporal consistency engine trust_scorer → trust_score + civic_urgency_score fusion math • Per-component file-role table (10 files explained) • Pipeline 5 and 6 flow diagrams added • Updated API examples with v2.1 response fields • USP table with problem statement alignment Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>1f1dab9 -
feat: v2.1 — trust scoring engine, temporal consistency, predictive hotspot, quick validate
Hikity +1: Claude Sonnet 4.6• NEW app/trust_scorer.py — unified trust_score (0-100) + civic_urgency_score (0-100) fusing deepfake (40%), scene (35%), routing (25%) signals across all pipelines • app/scene_analyzer.py — per-frame temporal consistency tracking (fraction of frames with detections), bbox coverage_pct, dominant_class, top_detection_confidence; 8 frames sampled (up from 6); actionable scene summaries for HIGH priority issues • app/router.py — enhanced Gemini prompt with sub_category + estimated_resolution_days; multi-keyword SCORED fallback (counts all hits, selects highest dept) replaces first-match; 40+ new Indian civic keywords across all 10 departments • app/main.py — 2 new endpoints: POST /predict-hotspot (Pipeline 5): geographic hotspot prediction via Gemini trend AI + statistical priority-weighted scoring fallback POST /validate-report (Pipeline 6): lightweight single-frame civic image validator All 4 existing endpoints now compute + return trust_score + civic_urgency_scor…8c9cf13 -
feat: add Pipeline 3 scene detection (YOLO road damage + TrashNet) and DistilBERT priority validator for dual-model consensus routing
Hikity39f212a
-
feat: implement pipeline 2 department routing using Gemini / keyword fallback
Hikity16783ba
-
fix: replace libgl1-mesa-glx with libgl1 for Debian Trixie compatibility
Hikity121be9d -
fix: Python 3.10 base image, CPU PyTorch, Linux temp path, HF port 7860
Hikity0bcc1e7