eerabhatt/ao-damage-estimation

Image Upload
    │
    ▼
Preprocessing
    ├── Resolution cap (800px max — CPU inference optimization)
    ├── Orientation correction (EXIF)
    ├── Quality gate (blur, brightness, resolution)
    └── Fraud signals — flagged in API response and audit log
            ├── low_pixel_variance: nearly uniform image — suggests a solid fill
            │   or digitally generated image rather than a real damage photo
            ├── editing_software_detected: EXIF Software tag contains Photoshop,
            │   GIMP, Lightroom, etc. — image was manipulated after capture
            └── duplicate_image: perceptual hash matches a photo submitted within
                the last 60 seconds — same damage being claimed more than once
    │
    ├──▶ Mask R-CNN (parts) ──▶ Part detections (class, bbox, mask, score)
    │
    ├──▶ YOLOv8m (damage) ──▶ Damage detections (class, bbox, score)
    │
    ▼
IoU / Mask Overlap Cross-Reference
    └── YOLOv8n-cls (severity) ──▶ Minor / Moderate / Severe per damage crop
    │
    ▼
Cost Estimation (GradientBoosting ML model)
    ├── Repair/replace decision per part (glass/severe damage → replace, else repair)
    ├── Regional labor rates by state — body / mechanical / paint (SCRS 2024 survey data)
    └── Total cost range (±15% band)
    │
    ▼
Gemini 2.5 Flash — natural language explanation + STP eligibility decision
    ├── STP criteria: cost < $1,500, confidence > 60%
    └── Auto-escalation to adjuster if confidence < 40%
    │
    ▼
Audit Trail (JSONL) — claim ID, session ID, timestamp, model version, full decision log

docker build -t ao-damage-estimation .
docker run -p 7860:7860 -e GEMINI_API_KEY=your_key_here ao-damage-estimation

cd backend
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install fastapi "uvicorn[standard]" python-multipart pillow numpy \
    opencv-python-headless pycocotools ultralytics scikit-learn joblib \
    google-genai python-dotenv torch torchvision \
    --index-url https://download.pytorch.org/whl/cpu
uvicorn api:app --reload --port 8000

GEMINI_API_KEY=your_key_here

cd frontend
npm install
npm start

git push space main

ao-damage-estimation/
├── backend/
│   ├── api.py              # FastAPI REST API — /detect, /estimate, /claims endpoints
│   ├── cv_detector.py      # CV wrapper: runs Mask R-CNN + YOLO, cross-references results
│   ├── cost_estimator.py   # ML cost estimation (GradientBoosting) with state labor rate adjustment
│   ├── llm_client.py       # Gemini integration: explanation generation + STP decision
│   ├── audit_logger.py     # JSONL audit trail — one record per claim
│   ├── fraud_detector.py   # Perceptual hash duplicate detection + EXIF metadata anomaly detection
│   ├── data/
│   │   ├── repair_costs.csv    # Part repair/replace costs
│   │   └── labor_rates.csv     # Body/mechanical/paint rates by state (SCRS 2024)
│   └── mask_rcnn/
│       ├── config.py       # Hyperparameters, class labels, paths
│       ├── model.py        # Mask R-CNN model factory
│       ├── dataset.py      # Custom PyTorch Dataset with COCO-format annotations
│       ├── train.py        # Two-phase training loop with AMP and gradient checkpointing
│       ├── inference.py    # Full inference pipeline with mask overlap cross-referencing
│       ├── evaluate.py     # COCO mAP evaluation
│       └── preprocess.py   # Image quality gating, orientation correction, fraud signals
├── frontend/
│   └── src/
│       └── components/
│           ├── ImageOverlay.js     # HTML5 Canvas bbox overlay, color-coded by severity
│           ├── ResultsDisplay.js   # Full results panel — STP, cost, state selector, overrides, history
│           ├── ImageUpload.js      # Drag-and-drop image uploader
│           └── LoadingOverlay.js   # Analysis loading state with cancel button
├── download_models.py      # Downloads model weights from HF Hub at build time
├── models/                 # Model weights (not in git — hosted on HF Hub)
├── audit_log.jsonl         # Append-only claim audit log (not in git)
└── Dockerfile              # Containerized deployment for HF Spaces

{
  "detections": [
    {
      "part": "Front-bumper",
      "damage_type": "Dent",
      "confidence": 0.73,
      "severity": "moderate",
      "bbox": [120, 340, 450, 520],
      "iou": 0.42
    }
  ],
  "cost": {
    "damaged_parts": [
      {
        "part": "Front-bumper",
        "damage_type": "Dent",
        "severity": "moderate",
        "action": "repair",
        "labor_category": "body",
        "labor_rate": 67.0,
        "cost_range": [373, 505]
      }
    ],
    "total_cost_range": [373, 505],
    "state": "MI",
    "labor_rates": {"body": 67.0, "mechanical": 95.0, "paint": 65.0}
  },
  "explanation": "The vehicle sustained a moderate dent to the front bumper requiring repair...",
  "confidence_score": 0.81,
  "stp_eligible": true,
  "stp_reasoning": "Claim eligible for auto-approval: cost $439 under $1,500 threshold, 81% confidence meets requirement.",
  "requires_adjuster_review": false,
  "override_allowed": true,
  "model_version": "1.0.0",
  "fraud_flags": [],
  "claim_id": "d4d22393-32b5-4720-afdd-44977b980943",
  "state": "MI",
  "inference_ms": 842.3
}