Long-form engineering write-ups: the problem, the approach, the architecture, and the measured outcomes behind real LLM systems.
Designing a low-latency, cost-controlled retrieval-augmented generation API on Bedrock and OpenSearch — chunking, hybrid retrieval, evaluation, and the caching that made it affordable.
When vector search kept failing on 'how are these connected' questions, I built a hybrid retriever over a knowledge graph and an embedding index — and let the query decide which one to trust.
A multi-agent copilot that reads sustainability disclosures, scores them against a rubric, and routes borderline calls to human reviewers — built for auditability, not just accuracy.