01 hi, I’m Roy 👋
Software engineer + ML researcher.
Currently monetizing products @ Databricks.
02 things I’ve made
A few systems, experiments, and side quests — from multi-node RL training to the tiny details of a good product interaction.
RL post-training for code reasoning
A fully open-source coding model trained with distributed reinforcement learning. I worked on the RLLM pipeline, asynchronous infrastructure, execution-grounded evaluation, and the reward loop that makes “good code” testable.
Read the research ↗Autonomous applications
An execution engine for LLM actions, designed around post-facto validation, undo, and damage confinement — so agents can do more while humans stay in control.
See the paper ↗AI sensor fusion
A vision-transformer pipeline for methane-leak detection from infrared camera feeds, developed with Squishy Robotics and the Berkeley Expert System Technologies Lab.
Read the IMECE paper ↗03 the public side of my brain
A small tour through the public repos: serious research infrastructure, useful little tools, and a few projects that started with “how hard could it be?”
Agentic RL on any harness, with any backend, on any benchmark. The open-source framework behind a lot of the “make agents learn” work.
Making the search for a place on Airbnb a little less painful. A small, scrappy automation project with a very relatable motivation.
A fine-tuning tool for Gorilla, connecting practical model adaptation with the API ecosystem that made the project fun in the first place.
A Discord bot for live COVID statistics and quick data visualizations. One of the earlier projects where data became something people could actually use.
04 selected papers
Distributed reinforcement learning for an open 14B code-reasoning model.
An execution runtime for LLM actions built around undo, validation, and damage confinement.
AI sensor fusion for detecting and characterizing methane leaks in difficult field environments.
05 a very short autobiography
I like moving between research and production, and I’m happiest when the feedback loop is short and the problem is a little weird.
Software Engineer
Machine Learning Researcher · DeepCoder-14B, distributed RL, evaluation
LLM runtimes, distributed systems, and cloud computing research
Methane detection, AI sensor fusion, and field robotics
B.A. Computer Science & Data Science · M.S. EECS
06 come say hi
I like sharp questions, research with an exit ramp, and projects that are a little too ambitious.
Connect on LinkedIn ↗