I build evaluations for robot policies.
The metrics and statistics that decide whether a policy is usable in the real world. I study math and computer science at Stanford, and this summer I built the policy evaluation pipeline at Parametric (YC F25).
The arm here is an animated Franka Panda acting out pick-and-place rollouts. Move your cursor over its table to choose where the next block starts. Since you arrived it has run 0 rollouts and 0 succeeded
Evaluation at Parametric
A policy that succeeds once might succeed 30% of the time. A real evaluation repeats the task many times, records every outcome, and reports an interval next to the success rate, so the number means what it says. This summer I built Parametric's internal evaluation pipeline for robot manipulation policies, and designed the metrics and statistical analyses used to decide whether a policy was ready for real-world use. I also rebuilt the robot telemetry for real-time remote debugging of policy runs, cutting chart payloads by over 100x.
The sixteen arms in view act out the same task, each with a true success rate of 74%, which the 95% interval should usually contain. Together they have logged 0 rollouts, and 0 succeeded
Calibration Precedes Evaluation
Written with John Alvarez for the Parametric blog, July 2026. If a task is too easy every policy passes, and if it is too hard every policy fails. We raise the difficulty one level at a time until the policies separate, then spend the rollout budget there.
The raised tables in view act out a toy version, with true success rates of 95%, 74%, 47% and 20% from level 1 to level 4.
VLA Memory
The arm here acts out a simple memory task: a light comes on behind one of three squares and goes off before the arm moves, and the block has to end up on that square. VLA Memory, my CS 224R project, targets tasks like this. I trained a Qwen3-VL-4B planner with persistent keyframe memory using streaming GRPO, driving a frozen π0.5 action expert. It reached 31.4% ± 3.1% success on the RoboMME ButtonUnmaskSwap benchmark, 10.1 points above the MemER-IL baseline, with zero human demonstrations.
In this animation the right square comes up 0 times in 0 tries, against a set rate of 80%.
Scam-victim protection eval
When someone who is being scammed asks a chatbot for help, does the model recognize the scam and keep warning them after they push back? I'm building a multi-turn evaluation for it, graded against six bright-line rules taken from FTC guidance, with two judges that must both quote the words that break a rule.
Other work
- 2026MetaChat 2.0 evaluation set At Stanford's Fan Lab I built the LLM evaluation set for MetaChat 2.0, a multi-agent framework for autonomous nanophotonic device design, and contributed to the framework itself.
- 2026Everen Co-founder. A local-AI app for private clinical transcription and note generation: Tauri desktop app, Rust backend, encrypted local storage. We tested it with 10 therapy offices, and 8 joined the waitlist with stated intent to purchase.
- 2026CodeSentry Reviews AI-written code. Claude's flags only survive when a Semgrep rule backs them up.
- 2026CoInvest A multi-agent Claude pipeline that grounds each investment-thesis report in live market and news data, and lists the conditions that would disprove the thesis.
- 2026Latent rectified-flow CT Low-dose CT denoising in 5 to 10 steps with a per-pixel uncertainty map. CS 231N.
- 2025Reproducible RL pipeline PPO with seeded determinism; two runs of one config match bit for bit.
About
Math major with a CS minor at Stanford, class of 2028. Xfund Ethics Fellow, Startups to Join research fellow, and a member of the Blythe Fund and the Stanford AI Club. I spent summer 2026 at Parametric; before that I was a research assistant in Stanford's Fan Lab and an engineering intern at Platform Science. Outside work I lift, golf, travel, and visit family in Brazil whenever I can.
While you read, the animated arms on this page acted out 0 rollouts.
The arms are an animation: each rollout's outcome is a random draw against a success rate set in the page's code, and the motion is scripted to match. The t-shirt folding chart uses real results from the Parametric post. Robot model: Franka Emika Panda from MuJoCo Menagerie (Apache-2.0), simplified for the web.