Evincta
Flagship AI system. A claim — damage photos, a scanned policy, and adjuster notes — is extracted by vision, matched against visually- and textually-similar past claims in a multimodal vector database, and reasoned over by a panel of specialised agents coordinating across four real Model Context Protocol servers. Deterministic tools decide the facts; a single constrained LLM synthesises the decision. Every recommendation passes a structural human-in-the-loop gate and is written to a tamper-evident audit log. Deployed on Vercel + Render, hardened, and gated in CI by a held-out evaluation.
The problem
Auto-claims triage is 20–30 minutes of multimodal context-gathering per claim — read the damage photos, check coverage, estimate cost, screen for fraud, find precedent — before any decision. It's slow, inconsistent between adjusters, and exactly the judgment-under-evidence work where an AI system should gather and synthesise while a human keeps the decision.
The solution
A LangGraph state machine with four deterministic specialist agents (coverage, cost, fraud, precedent) that call MCP tools for facts, and exactly one LLM synthesiser that turns those facts into a decision constrained to a typed schema — so a model hallucination can never produce an out-of-policy outcome. Coverage is an absolute deny-gate; the graph cannot commit a decision without a human approve/override, enforced by a named interrupt node. On any dependency failure it degrades to human review rather than deciding on incomplete evidence.
The outcome
Scored end-to-end on 18 held-out claims (never indexed) with asymmetric scoring and a CI gate that fails the build on regression: weighted decision accuracy 0.722 with no dangerous errors, payout-in-range 0.769, extraction 0.889/0.833. Hardened — retry · circuit breaker · rate limit · daily spend cap · input/output validation · structured logging — each verified by tests in CI. Traced end-to-end in Langfuse. The held-out eval caught five real bugs, fixed as clean commits.
Stack
