Building AI that checks its own work.
Ex-Google DeepMind, AWS & Microsoft. Stanford CS + ME.
I build systems that learn, then systems that check the learning. From reinforcement-learning racers to agents at Google DeepMind, the lesson is the same: the model is only as good as the proof behind its answer.
Today I'm co-founding Energent.ai (formerly CambioML, YC S23). One agent does the data work. A second agent, with fresh context, re-derives every figure and traces it back to the source before anything ships.
Before that: Google DeepMind, AWS and Microsoft, plus dual master's degrees in Computer Science and Mechanical Engineering at Stanford. Motto, unchanged since day one: learn everything.
AI that audits AI. An analyst agent handles the data work, and an auditor agent re-counts rows, re-derives figures and returns pass/fail with an evidence trail.
Worked on AI at Google DeepMind, then open-sourced the map out: xg2xg-ai, a guide to Google's AI/ML tooling and its open-source equivalents.
Reinforcement learning and simulation: DeepRacer environments, ROS bridges and distributed simulation tooling for training autonomous racers.
Software engineering at Microsoft. The first lap: learning how real products ship.
Dual master's: Computer Science and Mechanical Engineering. Where the bits met the atoms, including CS236 deep generative models and ME218B mechatronics, where he built DaleBot.
Stanford ME218B · Team Dale · Autonomous robot challenge
Before agents, a robot. DaleBot was a fully autonomous 12-inch vehicle built from scratch: it navigated an 8'×8' field to capture "polling stations" while a flywheel shooter launched foam balls at target buckets. Laser-cut hardboard and acrylic, hand-wired boards, and embedded C on top.
Full build log ↗
Energent.ai accuracy on the DABstep data-analysis benchmark (Hugging Face)
Stanford master's degrees, CS + Mechanical Engineering
Big-tech AI teams: DeepMind, AWS, Microsoft
Y Combinator batch, Summer 2023