Aptino

Summary

A multi-agent investigation platform that maps health-insurance claim facts to policy requirements through a structured evidence matrix, so every decision is grounded in retrieved policy text.

The problem

Claim adjudication depends on reading a large policy document and matching each fact in a claim to the clause that governs it. Doing that by hand is slow, and a summary that loses its citations cannot be defended in an audit.

What I built

Treat the system as investigation rather than prediction. Agents normalise the claim, plan what evidence is missing, retrieve the governing clauses, build an explicit evidence matrix, then reason over it under a fail-closed rule: when evidence is missing or conflicting, the system abstains and returns NEEDS_REVIEW instead of guessing.

Architecture

• 01Case Analysis Agent normalises facts and flags conflicts and missing evidence
• 02Hybrid retrieval: dense vectors and BM25 fused with Reciprocal Rank Fusion
• 03Post-fusion reranker narrows to high-precision policy chunks
• 04Policy Evidence Agent builds the evidence matrix with typed findings
• 05Decision Agent applies a safety hierarchy: Missing → Exclusion → Limits → Admissible
• 06Validation Agent checks every material claim against a retrieved citation
• 07FastAPI endpoints (/analyze, /review, /health) behind a Streamlit reviewer workspace
• 08Full provenance chain recorded as audit events for every analysis

My role

Built end to end — multi-agent orchestration, retrieval pipeline, evidence and decision modelling, API layer, reviewer UI, evaluation harness, and documentation.