Hackathon2026

CLAUSE

An insurance-claims adjudicator that returns APPROVE, FLAG, or DENY decisions grounded in verbatim policy clauses, with citations programmatically verified before display.

FastAPIGemmavLLMROCmAMD MI300XFireworks AI

CLAUSE was built for the AMD Developer Hackathon: Act II as an auditable insurance-claims adjudicator. It produces an APPROVE, FLAG, or DENY decision, but the decision is only useful if the evidence can be trusted.

Evidence that fails closed

Every displayed citation is programmatically checked against the source policy. If the quoted clause does not exist verbatim, the evidence is visibly rejected rather than silently dropped or replaced with a plausible-looking model output.

One structured model call

The adjudication uses a single structured-output Gemma call rather than a chain of agents. That keeps latency, failure modes, and audit reasoning easier to understand while still producing a typed decision, rationale, and supporting clauses.

Portable model serving

One environment variable switches inference between Fireworks AI and self-hosted Gemma served with vLLM on ROCm-backed AMD Instinct MI300X hardware. The GPU endpoint is rate-limited for operational safety.

Audit records

Each adjudication produces a SHA-256-backed audit record, tying the rendered decision to its inputs and evidence trail.

Designed & built withby Dhevenddra

Built with Next.js, Tailwind CSS & Framer Motion