Challenge · 01
Muzzle photos vary wildly in the field — mud, wet noses, low light, sideways angles.
Resolution
Heavy field-condition augmentation in training, plus an on-device quality gate that prompts a re-shoot instead of poisoning the match.
← All case files
Case / 2024.09 · Insurance · Agritech
Hide Nothing
Cattle-claim insurance, biometric identity
CV-driven cattle identification and automated claim filing — fraudulent claims caught at intake, before any payout.
/Insurance-Carrier/ · Shipped
91% of frauds caught at intake · clean-claim turnaround from 4–8 weeks to under 48 hours.
/ 01 · Stakes
What was on the line.
A regional carrier insuring smallholder cattle herds was paying low-eight-figures annually on fraudulent death claims. The pattern was crude but effective: the same animal re-enrolled under a neighbour’s policy, then “died” twice.
Field assessors couldn’t be everywhere, and a photo by itself didn’t prove much — two black-and-white cows look like two black-and-white cows. Legitimate claims took 4 to 8 weeks of paperwork; the fraud cases looked indistinguishable until after the wire transfer.
/ 02 · Approach
How we built it, in five moves.
- 01
Biometric enrollment at policy bind
The rancher’s phone captures profile, muzzle and head-side frames. Three on-device encoders — muzzle print, anatomy (eyes, ears, horns) and body markings — produce a signed embedding queued for upload.
- 02
Field capture at claim time
The same app captures fresh imagery, geotagged and EXIF-checked for tampering. A quality gate forces a re-shoot; nothing leaves the device unsigned, and captures queue offline.
- 03
Identity match before payout logic
Fresh embeddings are matched against the policy’s enrolled animal with an adaptive per-region threshold. A failed match routes to a fraud investigator before any payout decision is computed.
- 04
Cause classification + policy reasoning
Verified claims get a second pass to classify cause of death against exclusions and coverage windows. The engine recommends; a human approves, holds or overrides.
- 05
Audit-grade evidence trail
Every step writes a signed record into a hash-chained ledger the regulator can export directly — a verifiable chain of custody from first photo to final payout.
/ 03 · Challenges
What got in the way, and how we resolved it.
Challenge · 02
Breed variance swamped a single encoder; what worked in one region drifted in the next.
Resolution
Per-breed adapter heads on a shared backbone — a new region triggers an adapter fine-tune, not a full retrain.
Challenge · 03
Rural connectivity is often nonexistent.
Resolution
Capture, sign and queue fully offline; upload opportunistically; a daily reconciliation report flags missing uploads.
/ 04 · Outcomes
What changed, in numbers.
- 0
- Frauds paid post-deploy
- <48h
- Claim TAT, clean cases
- 91%
- Frauds caught at intake
- 12pp
- Loss-ratio improvement
- Fraudulent claims caught at intake, before any payout
- Legitimate claim turnaround compressed from 4–8 weeks to under 48 hours for clean cases
- Regulator-grade evidence surfaced by the audit trail with no extra work
- Loss ratios improved by double-digit percentage points within two underwriting cycles
/ 05 · Evidence
Live across multi-region farms on two underwriting books. Continuous model retraining as new breeds enroll. Audited internally and by the client’s GRC team annually.
- Flutter mobile capture
- ONNX Runtime on edge
- Milvus vector store
- Python + Ray
- PostgreSQL append-only ledger
