ClaimSense
On-device AI motor insurance platform — fuses computer vision and sensor data to score driver risk in real time, even offline
The insurance assessment gap
In India, motor insurance claim assessment is almost entirely manual — a surveyor visits the site, photographs the damage, and files a report 3–7 days later. The process is subjective and leaves room for inflated claims. ClaimSense was built to bring on-device AI to the point of claim: a field agent films the damage, gets a risk-scored report in under 10 seconds, and the model runs entirely offline.
Context
- Must work offline — rural breakdown locations have no signal
- Inference must be real-time — surveyors can't wait 30 seconds per frame
- 6 damage categories: bodywork, glass, tyre, structural, interior, underbody
- Validated against real vehicle damage in live field trials
Our approach
Key decisions
YOLOv8 fine-tuned on motor damage
Fine-tuned on a labeled motor damage dataset. Detects and classifies 6 damage categories simultaneously in a single forward pass. 94.2% mAP on held-out validation set.
ONNX Runtime for on-device inference
PyTorch model exported to ONNX and INT8 quantized. Runs at under 100ms per frame on a mid-range Android device with no GPU — no cloud call, no latency.
IMU sensor fusion for driver risk
Accelerometer and gyroscope data feed a separate lightweight risk model. Hard braking, sharp cornering, and high-speed events generate a driver risk score independent of camera input.
Firebase for offline-first claim sync
Reports are written locally first and synced to Firebase when connectivity is restored. No claim is lost due to poor signal in a rural area.
Results
What we achieved
mAP on damage detection
Inference per frame on Android
Damage categories classified
Validated in vehicle field trials
Stack used
Next project
Jain Poddar & Co. →Start a project
Let's build something great
Drop your details below. We'll have it open in Gmail, ready to send — no copy-pasting.