Putting agents to work on real codebases: specifications, workflows, evaluations, observability and memory. With the Sutro Group, I built the research workspace and ran agents against whole catalogs of experiments. I build the infrastructure for understanding what those agents did and whether their work holds up.
Hinton Problems and Schmidhuber Problems 2026, with Yaroslav Bulatov and the Sutro Group
With the Sutro Group, I ran coding agents that reimplemented 53 of Hinton's experiments and 58 of Schmidhuber's papers in pure NumPy.
At Roby, I was Head of ML, working on reinforcement learning for HVAC optimization. At On Deck, I was a founding engineer and later a senior data engineer, working on data infrastructure and matchmaking algorithms for founders.
My Pixel Pilot and KayOS
2024 to 2026
I started My Pixel Pilot to manage my own work: task and project management with LLMs, imports from Jira, Trello
and Monday, custom workflows and a task dispatcher. I built it around self-hosting and bringing your own models,
including local models.
I was building it when I co-founded KayOS, where I served as CTO. The company’s work spanned production AI
platforms for legal tech, marketing analytics and business intelligence. My work included the infrastructure
beneath them—knowledge graphs, AI sandboxes and ETL frameworks—and an AI coding IDE.
Head of ML and Data at the NEAR Foundation. NEAR Atlas was the foundation’s analytics product over the chain’s data, 770M+ transactions, and the pipelines under it. I published the NEAR Protocol public dataset on Google BigQuery, the dataset NEAR’s documentation points people to today.
Penetration testing, application security and LLM red teaming, including prompt injection, excessive agency and unsafe output handling.
Pentesting and LLM red teaming
2015 to now
Penetration testing, LLM red teaming and application security, since the CongruentAI years. In 2025 my team, unit420, competed in OpenAI’s gpt-oss-20b red-teaming competition on Kaggle: 158 attack prompts across four categories, a 27.2% success rate, and a hybrid refusal detector built first, because the keyword baseline flagged proper refusals as hits 73% of the time. I wrote a pentester’s guide to the OWASP Top 10 for LLM applications. Stanford Advanced Cybersecurity Program (2023), CompTIA Security+ and Network+.
Machine learning, linear algebra and working with coding agents. I've taught Georgia Tech OMSCS students through Codementor, run a deep learning study group in San Francisco, and publish lessons at AgenticEngineer.ING.
CongruentAI, founder, 2015 to now. CEO from 2015 to 2020: ML consulting, applied research, security work and open source. Since 2026, my research on coding agents runs under it.
KayOS, co-founder and CTO, 2024 to 2026. 13 production AI platforms, the agent infrastructure under them, an AI coding IDE.
NEAR Foundation, Head of ML and Data, NEAR Atlas, 2022 to 2024. The NEAR Protocol public dataset on BigQuery.
On Deck, Senior Data Engineer, 2020 to 2022. Founding Engineer, 2019 to 2020. Data infrastructure and matchmaking algorithms for founders.
South Park Commons, independent machine learning research, 2019 to 2020. Program synthesis, and Mora, a search engine that turned AI research papers into runnable code.
Roby (Alexa Accelerator, powered by Techstars '18), Head of ML, 2018 to 2019. Reinforcement learning for HVAC systems.
Codementor, Lecturer in Computer Science (AI/ML), 2015 to 2016. Machine learning and linear algebra for Georgia Tech OMSCS students.
DraperTV, San Francisco, Full Stack Engineer, then Technical Team Lead, 2013 to 2015.
Also
rLoop, Reddit's Hyperloop team SpaceX Hyperloop Pod Competition
Simulation models for rLoop, the Hyperloop team that formed on Reddit and reached the SpaceX competition finals.
Mora 2019 to 2020, with South Park Commons
A search engine that turned AI research papers into runnable code.
