Real work doesn't live in one file type. It lives across CAD geometries, engineering drawings, PDFs, tables, and documents that were never meant to work together. I build representations that unify them and extraction systems that keep every claim traceable — so the patterns hidden across the corpus become visible.
You're sitting on information that spans CAD files, drawings, and piles of documents.
No single view connects them, and no answer can be trusted without a source.
Representation is fragmented
CAD geometries, engineering PDFs, tables, and text each need different models. Without a unified representation, connections across modalities never surface.
Extraction lacks grounding
Summaries and chat answers are easy to generate, hard to defend. Without provenance to the exact page, geometry, or document, they can't be trusted for real decisions.
Patterns stay hidden
The most important signals only emerge across hundreds of files — contradictions, outliers, recurring structures. No one file reveals them alone.
I focus on the full loop: how information is represented, how it is extracted, and how emergent patterns are surfaced — always with traceability to the source.
Representation Across Modalities
Modeling CAD files, engineering drawings, PDFs, and documents into a shared, structured representation — so geometry, text, tables, and images can be queried together.
How you represent data determines what you can ask. Get that right, and the corpus stops being a pile of files and starts being a system you can reason about.
Extraction With Provenance
Extracting entities, facts, and relationships — each linked to the exact source: page, geometry, section, and confidence.
No black-box summaries. Every insight is inspectable, and every answer can be defended back to where it came from.
Emergent Pattern Discovery
Finding what only appears when you look across the whole corpus — recurring structures, contradictions, outliers, and subtle signals.
This is where representation pays off: patterns invisible in any single document become obvious in the right model.
Systems That Compound
Building so that judgment and structure accumulate — corrections and links applied forward, so the next thousand files are easier than the last.
Institutional knowledge should live in the representation itself, not just in someone's head.
The hardest information problems aren't text-only. They require reading geometry and documents together.
CAD Files, Drawings & Engineering PDFs
At Dirac I processed 1M+ geometries and unstructured engineering PDFs — learning where generic document tools break and where geometry-aware representation matters. A dimension in a CAD file and a note in a drawing have to resolve to the same thing.
Parse and link geometries, drawings, and structured outputs into a common representation
Surface contradictions: figures that don't match across sources, missing annotations, unresolved references
Trace every extraction back to its exact geometry, page, or document for defensibility
Build models that compound across projects, so what was learned once carries forward
"CAD and documents are not separate problems. The real structure only appears when you represent them together and let patterns emerge across the whole set."
Documents, Tables & Other Data
Beyond CAD: PDFs, spreadsheets, images, and unstructured documents — all part of the same representation problem. The goal is always one queryable view, grounded to source, where emergent structure is visible.
Extract observations across file types with consistent structure and source citations
Link entities and facts across documents — people, terms, obligations, references
Query across hundreds of files as easily as one, with provenance intact
Reuse structure: what was modeled once doesn't need to be rebuilt from scratch
Model the data, not just the files
Start with representation. For CAD, that means geometries as first-class objects; for documents, it means logical sections and linked entities — not just chunks of text.
Extract with grounding
Every extraction is confidence-scored and traced to its exact location — page, geometry, or document. High confidence flows through; uncertainty is where human judgment matters.
Let patterns emerge
Once the corpus is well-represented, contradictions, outliers, and recurring signals surface naturally. The representation does the work that no amount of per-file prompting can.
Make insight durable
Structure should persist and compound. A correction or a new link should improve the next thousand queries — not disappear when the session ends.
I've built extraction, structured pipelines, and production ML across environments where information is multimodal, messy, and high-stakes — and watched the same gap appear: scattered inputs, no shared representation, answers you can't defend.
Meta (2022–2023)
Built ML systems serving 50M+ users daily and improved hate-organization detection by 15% PR-AUC. The lesson: production ML lives or dies on representation. An output only matters when it reflects what the system faithfully models and can trace back to its source.
Dirac (2023–2025) — Head of AI
As Head of AI, I processed 1M+ geometries and unstructured engineering PDFs, then built confidence scoring and structured pipelines from scratch. Reduced latency 70% and workflow interruptions 90% — by representing CAD and documents together instead of treating them as separate silos.
Microsoft (early career)
Early production software engineering at platform scale. It set the standard for everything that came after: ship reliable systems real users depend on, and make them explainable. That standard is why every extraction I build grounds itself in its source.
"I've learned to read across modalities — CAD, documents, and other data — and to build representations where emergent patterns finally become visible, and every answer traces back to where it came from. That is the work I am best positioned to do."
— Vatsal Trivedi
Representation and extraction are not side problems — they are the problem. When they are right, patterns compound and answers become defensible instead of starting over with every new file.