I build document intelligence and LLM systems that turn messy information into answers people can trust, with evaluation, access controls, and real workflows behind them.
A production RAG assistant for legal documents. OCR makes scanned PDFs retrievable; semantic search and caching reduce repeated calls. Multi-user sessions, role-based access, audit visibility, logging, and a labeled evaluation set support reliable use.
A routing layer that sends each LLM request to the cheapest model that can reliably handle it. Routing decisions come from a benchmark harness, not intuition: every prompt ran three times per model and was labelled with the cheapest model that passed every run. Deployed live with Docker and CI/CD, plus budget caps, latency fallbacks and per-request cost logging.
Built multi-format ingestion and retrieval for PDFs, images, and Excel files, connecting chunking, embeddings, and semantic search to backend chatbot workflows.
Built forecasting and reporting systems across supply chain and growth operations, from stock-level prediction to customer cohorts and media-efficiency models.
Building a RAG platform with OCR, semantic retrieval, caching, access controls, and a 500+ question evaluation framework.
Owned forecasting, analytics, AI knowledge workflows, and reporting that reduced manual work and supported growth decisions.
Built document ingestion and RAG components for an Azure-based enterprise AI platform.
Built inventory forecasting models and monitoring pipelines, helping reduce stockouts by 25%.
Built a quality monitoring platform for production defects, with searchable records, classification logic, and statistical reporting.