"We're here to put a dent in the universe."
Research engineer · product builder
Who do you think you’re looking at? When the model breaks, I don’t call someone else, Skyler. I’m the one they call. Agentic systems, LLM post-training, 2D and 3D perception. You see a demo run once on clean data and think that’s the job? No. I’m not the one watching the demo. I’m the one who ships.
I've always looked at the stars and wondered about our place in the universe. From both physical and spiritual perspectives, I keep returning to two questions: Who am I, and where do we go next as a civilization? I also love history, financial markets, geopolitics, and theology. I reset by running, hiking, or lifting; play football, cricket, and tennis; and unwind with sitcoms.
I'm also a firm believer in e/acc and in our ability to solve difficult problems when we set aside our differences and work toward becoming a Type I civilization on the Kardashev scale. AI and robotics will play a major role in that journey.
How I work
I move from an outcome to the uncertainty blocking it, test what matters, and use production evidence to decide what the next iteration should be.
01 / Guiding question
In practice
For document intelligence, better OCR was not enough. The target was 100% decision accuracy: when the system compared outputs, it had to identify the correct one and reject the incorrect one, with no ambiguous result.
02 / Guiding question
In practice
Handwriting, scan quality, layout, domain symbols, and schema rules became separate system constraints instead of one vague model problem.
03 / Guiding question
In practice
I benchmarked vision-language models, decomposed failure modes, and tested where OCR, generation, and verification each belonged.
04 / Guiding question
In practice
Human corrections feed the verification database, while production failures become reproducible cases for the evaluation harness.
Production evidence loops back into the next outcome. This is a cycle, not a handoff.
Agent eval harness design across multiple projects and 3D reasoning on outdoor imagery.
My favorite kind of problem
I close that gap for a living — evaluation, post-training, and the model work underneath it. The application layer is downstream of getting that part right.