Enterprise GRC and Compliance AI Platform
Core AI engineering on a governance, risk and compliance platform for an enterprise client. I built the retrieval and reasoning layer behind a multi-persona AI consultant system used for regulatory guidance, where a wrong answer is not an inconvenience but a compliance exposure.
Architecture: hybrid retrieval with permissions enforced at the retrieval layer
Engineering highlights
- Hybrid retrieval combining dense vector search with lexical search plus a cross-encoder reranking stage, rather than vector similarity alone
- Three-tier multi-tenant permission model enforced inside the retrieval path
- Structure-aware chunking with precedence and freshness rules for weighting conflicting source documents
Evaluation work
- Benchmarked 16 candidate local models across persona adherence, retrieval quality and hallucination risk before selecting a production model, which surfaced one high-scoring model that was also the most dangerous on hallucination
Result
Core retrieval architecture shipped into a production compliance platform, with a sanitized write-up published publicly under written client authorization.
Under the hood
.NET backend · Angular frontend · Hybrid retrieval with reranking