I work on representation of data and extraction of information across multimodal data — CAD files, engineering drawings, PDFs and other documents — building systems that surface emergent patterns others miss.
Dirac
2023 - 2025 · Head of AI
Meta
2022 - 2023 · MLE
Microsoft
Early career
Information is rarely clean. It lives in CAD geometries, in layered engineering PDFs, in scattered documents and systems. I build representations that make that information queryable — and extraction systems that make it reliable — so emergent patterns become visible early.
Represent
Turn multimodal inputs — CAD files, drawings, tables, text, images — into structured, linked representations.
Extract
Extract entities, facts and relationships with grounding to the exact source page, geometry, or document.
Surface Patterns
Find what emerges across the corpus — contradictions, outliers, and signals invisible in any single file.
The throughline from Dirac to Meta to today is the same: messy, multimodal inputs → structured, source-linked representations → systems that let people ask hard questions and trust the answers.
Retrieval alone finds the closest passage. It doesn't tell you how documents connect, or what pattern only appears when you look across hundreds of them. Deep understanding comes from representation — how information is modeled, linked, and traced — not from chat summaries that reset every session.
"The value is not in finding a file. It is in modeling what the files collectively mean — and proving every claim back to its source."
Dirac
2023 - 2025 · Head of AI
Built AI systems from scratch handling 1M+ geometries and unstructured engineering PDFs. Reduced user-facing latency by 70% and workflow interruptions by 90%. Learned how to turn messy, multimodal inputs — especially CAD — into structured, source-linked systems that surface patterns across a large corpus.
Meta
2022 - 2023 · Machine Learning Engineer
Built ML systems serving 50M+ users daily and improved hate-organization detection by 15% PR-AUC. Saw what production ML demands: representations are only useful when they are grounded, and patterns only matter when you can trace them to their source.
Microsoft
Early career
Early production software engineering at one of the largest platforms in the world. The foundation for everything that came after: shipping reliable systems that real users depend on.
Representation shapes what you can see
How you model data determines which patterns can emerge. Get representation right and insight follows.
Every extraction traces to its source
No un-sourced assertions. Every claim chains back to the exact page, geometry, or document.
Patterns compound, not reset
Insight should accumulate across files and over time, instead of starting over every conversation.
Multimodal is the real world
CAD, PDFs, tables, and documents together — the system has to read them all and link them faithfully.
I've spent my career turning messy, multimodal inputs — from CAD to documents — into systems that make information trustworthy and patterns impossible to miss.