I work on representation of data and extraction of information across multimodal data — from CAD geometries and engineering PDFs to tables, images, and other documents. My focus is building representations that make messy information queryable, and extraction systems that make it trustworthy, so emergent patterns become visible.
Independent Work — Representation & Extraction
Focused on how information is represented and extracted across multimodal, messy inputs — CAD files, engineering drawings, documents, and other data — toward systems that surface emergent patterns and keep every claim grounded in its source.
Every extraction traces to its source by design: a page, a geometry, a timestamp, or a document — so answers are not just retrieved, but defensible.
Dirac · Head of AI
Head of AI at a CAD-automation startup. Processed 1M+ geometries and unstructured engineering PDFs, and built extraction, confidence scoring, and structured pipelines from scratch. Reduced user-facing latency by 70% and workflow interruptions by 90%.
Core lesson: CAD and documents together are a representation problem — get the representation right, and emergent structure across thousands of files becomes queryable.
Meta · Machine Learning Engineer
Built ML systems serving 50M+ users daily. Improved hate-organization detection by 15% (PR-AUC), directly impacting 11M+ profiles daily and reducing false negatives by 39%. Learned what production ML demands at scale, and why grounded, well-represented outputs matter more than clever models.
Microsoft
Early production software engineering at one of the world's largest platforms. Where I learned to ship reliable systems that real users depend on, self-funding my way through school along the way.
Stax · Founder
Founded Stax, growing to 400+ weekly active users across 4 colleges, supporting 15,000+ classes. Invested $10K and managed the entire product lifecycle. Learned invaluable lessons about product development, user acquisition, and market validation.
Georgia Tech
Earned a degree in Mechanical Engineering with a minor in Computer Science (Intelligence thread), 3.4 GPA. Financed my entire education through internships at Microsoft, Capital One, and Cardlytics while maintaining strong academic performance.
"Every technical operation I've worked in runs on information that is scattered across CAD files, documents, and systems — and nobody can fully query it. I focus on representing that information faithfully and extracting it with provenance, so the patterns hidden across the corpus finally surface — and every answer traces back to its source."
— Vatsal Trivedi
Technical Execution
- •AI/ML at Scale: Production ML systems serving 50M+ users, processing millions of items daily
- •Multimodal Representation: CAD geometries, engineering drawings, PDFs, and documents into structured, linked representations
- •Information Extraction: Provenance on every claim, with a hard grounding rule against un-sourced assertions
- •Full Stack: Python, TypeScript, React, Next.js, TensorFlow, PyTorch, AWS, GCP
Approach
- •0→1 Building: Founded Stax (400+ users), led multiple greenfield AI projects at Dirac
- •Resilience: Self-funded education from age 16, managed full lifecycle from idea to launch
- •Emergent Patterns: Best positioned to find what only appears when you model the whole corpus, not just one file
- •Cross-Modal Linking: Connecting CAD, documents, and structured data into one queryable view
Systems Thinking
- •Representation First: How you model data determines which patterns can emerge
- •Grounding: Every answer traced to its exact page, geometry, or document
- •Compounding Insight: Knowledge that accumulates across engagements instead of resetting
- •Production Hardening: Built to be reliable, explainable, and trustworthy at scale
Domain Expertise
- •Multimodal Intelligence: Ingestion across CAD, PDFs, images, and documents into one reconciled view
- •Provenance & Trust: Every answer sourced to its exact page, timestamp, or geometry
- •Enterprise AI: Deployed production ML at Meta (50M+ users), know what breaks at scale
- •Pattern Discovery: Surfacing contradictions, outliers, and emergent signals across large corpora
Interested in multimodal data representation, information extraction, or finding emergent patterns in complex information?