Multimodal Representation & Extraction
Turning messy, multimodal inputs — CAD files, engineering drawings, PDFs, and other documents — into structured, linked representations. Every extraction is grounded to its source, so patterns that only emerge across the corpus become visible and defensible.
Not just search — representation
Retrieval finds the closest passage. Representation maps how information connects and surfaces patterns across the corpus.
Most tools handle text. Real work spans CAD geometries, drawings, tables, and documents together.
Summaries reset every chat. Structured, source-linked representations compound over time.
Dirac
Head of AI. Built systems from scratch handling 1M+ geometries and unstructured engineering PDFs. Led multiple 0→1 projects that reduced user-facing latency by 70% and workflow interruptions by 90%. The work was fundamentally about representation: turning CAD and documents into a coherent, extractable model.
Meta
Built ML systems serving 50M+ daily active users. Improved hate-organization detection by 15%, directly impacting 11M+ profiles and reducing false negatives by 39%.
Key Insight
Production ML lives or dies on representation. A model output is only useful when the underlying representation is faithful, and every result traces back to what the system actually knows.
Microsoft
Early production software engineering at one of the largest platforms in the world. The foundation for everything that followed: shipping reliable systems that real users depend on, at scale.
Stax
Founded Stax, growing to 400+ weekly active users across 4 colleges, supporting 15,000+ classes with personalized recommendations. Invested $10K and managed the entire product lifecycle from ideation to launch.
Lessons Learned
Validated market needs, developed user-centric solutions, and drove rapid growth. The experience provided invaluable lessons about product development, user acquisition, and market validation that shape how I approach complex information problems.
From production ML at Meta and Microsoft to AI built from scratch at Dirac, I turn messy, multimodal inputs — CAD, documents, and more — into representations that surface emergent patterns.