Gabriel Cha
i like to think, explain, build.
agentic harnesses that reason, search the web, and synthesize information fascinate me. so does fine-tuning llms on a100s to push quality on narrow tasks.
Software engineer at Pax (YC S24). Previously an AI/ML engineer at Honeywell and Root Access (backed by AlleyCorp), and a data scientist at Edison.
Selected work
-
Accepted to ACM SIGCSE 2026
ContentGen: Improving LLM-Generated Educational Content for Data Science
We needed practice problems for the intro data science courses we taught at UCSD, so we built a JupyterLab tool that drafts questions from the instructor's own notebook, then tested which prompts actually made them better. Six instructors tried it and found it useful.
-
yc interview, ’25 Y Combinator interview, Summer 2025
Cloi: a local debugging agent for the terminal
Most AI debuggers send your stack traces and source code to someone else's server, which is a dealbreaker when your code or data is private. I wanted one that never leaves your laptop. Cloi reads the error, pulls in the parts of your codebase that matter, and suggests a small fix you can accept or reject — because most bugs just need the right context.
-
Funded by the UCSD HDSI Undergraduate Research Scholarship
Speech emotion recognition with CNNs, SVMs, and ViTs
Can a model hear how someone feels? Two teammates and I trained models on 12,000 voice clips, and our best one got 86% right. What stayed with me was how sure it was when it was wrong — calling anger sadness with high confidence. That's what pulled me toward making models easier to understand.
-
Used by 2,000+ students across 12 terms of DSC 80
Loan Applications: the pandas project I wrote for DSC 80
As a TA, I designed this project from zero to one: I picked real loan application data, shaped it into a story about how lenders decide who gets which interest rate, and wrote the questions that walk students through it. They clean messy job titles and dates, bin credit scores, and run a permutation test on personal statements. More than 2,000 students have learned the fundamentals of data science from it.
-
An N-Gram language model from first principles
I wanted to understand language models from the ground up, so I built the simplest one: guess the next word by counting what usually comes next. It sounds fluent two words at a time but wanders off over a full sentence, which made me appreciate how much today's models quietly keep track of.