I build GenAI systems that actually ship.

Turns messy handwritten geotechnical boring logs into clean, structured, reviewable data - ~1,000 docs/day for 2,000+ users at a US geotechnical engineering firm.

• Async FastAPI + React/TypeScript app on a multi-stage Azure OCR pipeline (quality gating → classification → extraction) - cut manual data entry 70% and operating costs 60%.
• 14-step deterministic OCR post-processing engine lifted extraction accuracy 68.8% → 86.5% on a hand-labeled ground-truth benchmark.
• Two-stage LLM critic/corrector (Llama 4, Claude Sonnet, GPT on one Azure AI Foundry endpoint) with schema guards against hallucinated fields.
• Role-based dashboards with Microsoft SSO and a version-controlled review UI with 16x zoom for verifying sub-millimeter handwriting.

A multi-jurisdiction regulatory knowledge base with a chatbot that cites its sources - highlighted right inside the PDFs.

• Hybrid RAG (BM25 + vector + semantic search on Azure AI Search) over clause-level legal chunks spanning 9 jurisdictions - replacing a manual process that required 70–80 regulatory experts.
• Intent-tuned dynamic top-K retrieval, two-pass per-jurisdiction search, and LLM-judged chunk selection - delivering 80–90% accurate answers.
• In-PDF citation viewer: every answer links to the exact page with the cited passage highlighted, plus a verbatim mode that shows legal text untouched by the LLM.
• Production guardrails: tiered prompt-injection detection, scope boundaries, citation-relevance filtering, and per-request LLM cost tracking.

An LLM web app that does the boring parts of MicroStrategy → Power BI migration for you.

• Cut migration effort 40–45% - the LLM reverse-engineers data models from live MicroStrategy report SQL and batch-converts metric expressions to DAX.
• One-click Power BI dataset creation, with LLM-generated schemas grounded against the live warehouse before push - no hallucinated columns.
• BRD-to-schema generator: PDF/DOCX specs → normalized schema + DDL for 5 SQL dialects, validated by executing in a throwaway schema.
• Synthetic data generator sized for 1B rows (~1M rows in 10–20s, vectorized NumPy) with S3 → Redshift loading.