Vinícius Raposo

Applied AI Reliability Engineer

AI systems you can leave on without babysitting.

Profile pin · TypeScript

WCP-Compliance-Agent-V5

Five-service WH-347 payroll compliance platform: deterministic validation decides, the LLM explains, and unverified wage coverage goes to human review instead of approval.

377 tests, including evals over 100 golden examples: Contracts 42 · Agent 70 · Gateway 45 · Web 30 · Compliance Core 89 unit + 101 eval, plus four tooling checks; checksum-verified evidence.

Language: TypeScript Updated: 2026-08-15 Visibility: public
View on GitHub Back to catalog

Without deterministic validation owning the decision, WH-347 compliance becomes a black-box LLM guess—auditors cannot trust outcomes, retries drift, and violations slip through as plausible prose.

Deterministic validation decides; the LLM explains.

Offline demo:

pnpm demo:ui

Walkthrough

Deterministic checks rejected this WH-347 — violations, trust score, and citations; the LLM only wrote the explanation.
Deterministic checks rejected this WH-347 — violations, trust score, and citations; the LLM only wrote the explanation. Sample data.
The review queue surfaces low-trust cases flagged by deterministic validation.
The review queue surfaces low-trust cases flagged by deterministic validation. Sample data.
Analytics aggregate trust and volume for the same demo case across routes.
Analytics aggregate trust and volume for the same demo case across routes. Sample data.

Test distribution — 377 tests across five services

Contracts42Agent70Gateway45Web30Compliance Core190

The per-service split of the 377 tests cited above: Contracts 42 · Agent 70 · Gateway 45 · Web 30 · Compliance Core 190 (89 unit + 101 eval over the golden examples), plus four tooling checks and the SHA-256 checksum-verified evidence bundle.

Why it matters

Five-service WH-347 payroll compliance platform: deterministic validation decides, the LLM explains, and unverified wage coverage goes to human review instead of approval.

Reproducible proof

A credential-free scenario exercises extraction, validation, trust scoring, Agent synthesis, persistence and audit contracts, cache fallback, SSE replay, cost/latency metadata, and Web fixtures. The normalized bundle is verified with SHA-256 checksums.

Honest boundary

The canonical demo is local and offline. PostgreSQL, Redis, Docker, SAM.gov, Langfuse, and real LLM providers are optional integration surfaces; this page does not claim a hosted production deployment.

Repo facts

Default branch: main. Created: 2026-05-06. MIT licensed.