Rohan Thomas · Austin, TX · taking new projects
I spent six years shipping systems where mistakes are expensive: portfolio tools on $3T+ in assets at Charles Schwab, 50K packages a day at FedEx. Now I work on AI engineering standards across five U.S. Treasury bureaus. I hold every build to that bar.
Selected builds. Every one of these went from an idea to something running in production.
A healthcare agent built on OpenEMR's FHIR API, with 14 tools, a 6-layer verification pipeline, and a 92-case eval suite. Clinical staff can't act on answers that are only usually right, so every answer gets checked.
TTB label review is usually done by eye. This pipeline reads a label image and returns a field-by-field compliance verdict in seconds, with a violation report you can act on.
Polymarket and Kalshi often price the same event, worded two different ways. TruthLayer matches contracts with embeddings plus Claude verification, streams fee-adjusted spreads, and grades every call against real settlement outcomes — the scoreboard is proof, not theory.
An autonomous coding agent built on LangGraph. It plans typed edit steps, makes anchor-based file edits with fuzzy fallback, fans work out to parallel sub-agents, and routes each task to the right model tier. Most of what I know about agents came from building it.
A plugin runtime built into the open-source Chatbox client. Plugins live in iframes behind a postMessage lifecycle protocol and register as tools the AI can call mid-chat — chess (Stockfish WASM), weather, and Spotify prove the model works.
Online players get HUD stats; live players get a notebook. PokerStats logs a hand in under three seconds and turns your history into VPIP, c-bet%, and a leak report — every stat formula verified against worked examples in a 121-test suite.
Scores Reddit, YouTube, and news sentiment with Claude and blends it into one hype signal from −100 to +100. Sources fail independently, so one API outage degrades the index instead of breaking it. Rebuilt from a HackUTD project that won Goldman Sachs' investment-strategy challenge.
Three ways to work with me. Every engagement ends with something running in production.
A scoped build that goes from prototype to production, with an eval suite and a real deployment at the end.
I join your team and build the agent, eval, and interface work alongside your engineers. Short loops, working software every week.
Architecture and eval review for AI features you already have in flight. The goal is to find the failure modes before your users do.
I'm Rohan, an engineer in Austin. I build AI products for domains where wrong answers cost money. Right now that's the U.S. Treasury, where I work on AI engineering standards across five bureaus, including the IRS and TTB. Before that I spent four years on portfolio tools at Charles Schwab, where a bad deploy touches $3T+ in assets. I like small teams, short loops, and working software every week.
Now: building AI agents · taking on new projects