F.I.R.M

Description

F.I.R.M (Fake Interaction Recognition Model) is a lightweight, multi-agent system that evaluates screenshots and proofs uploaded by couriers and returns a per-proof score and category (FAKE, INCONCLUSIVE, SUFFICIENT, BONUS_WORTHY) to drive in-app feedback and operational workflows.

Objective

Automatically detect tampered or low-quality proofs so legitimate deliveries pass quickly while suspicious submissions are escalated for review.

Advantages

• Minimal infra: runs on CPU-friendly agents (heuristics + lightweight models), no mandatory GPU.

• Cheap deployment: works with existing storage/webhook infrastructure and can be run as small microservices or serverless functions.

• Low operating cost: batching, caching, and feature hashing reduce compute and storage footprint.

• Easy integration: simple API/webhook contract.

High success rate

Proven effective on common fraud patterns (reused images, basic tampering, GPS mismatches); tuned ensemble + human-in-loop keeps false positives low while preserving coverage.

Operational benefits

• Fewer manual reviews and faster dispute resolution.

• Reduced chargebacks and improved revenue protection.

• Faster courier feedback and higher quality submissions (fewer order retakes).

• Measurable reduction in fraud-related operational overhead and clearer audit trails.