Parametric · YC F25
Software Engineering Intern
Building out the policy eval pipeline, and designing the metrics and statistics that decipher if a policy is usable in the real world.
Lucas Burgett
Math + CS @ Stanford · Intern @ Parametric (F25)
I'm studying math and computer science at Stanford. I'm super interested in physical AI systems. I love learning about VLA's, Reinforcement learning, and how to deploy reliable policies in the real world.
Currently
Stanford research
Fan Lab, Stanford University
Contributing to the development of MetaChat 2.0, a multi-agentic framework for autonomous nanophotonic device design.
Education
Stanford University
B.S. Mathematics · Minor in Computer Science
Projects & Personal Research
VLA Memory
RL-trained memory for long-horizon robot manipulation
A hierarchical VLM + VLA system for robots that have to remember across tasks. A Qwen2.5-VL-7B planner is fine-tuned with GRPO to pick the keyframes that actually matter and issue the next subtask; a frozen π₀.₅ policy carries out the low-level motion.
Coinvest
An AI investment-thesis research tool: type a thesis, get back a structured report with data-driven confidence scores, company comparisons, historical parallels, and the conditions that would prove it wrong. A multi-agent Claude pipeline grounds every score in live yfinance and NewsAPI data, with scheduled re-research, portfolio tracking, and a daily morning brief.
Latent Rectified-Flow CT Reconstruction
Denoising low-dose medical scans with a latent flow model
Low-dose CT and X-ray scans are noisy, and standard diffusion models need hundreds of steps to clean them. This model learns a straight-line flow between paired low-dose and normal-dose images in a compressed latent space, so a scan reconstructs in 5 to 10 solver steps.
- A frozen VAE compresses each scan; a learned vector field transports the noisy latent toward the clean one along a straight path, trained by velocity matching.
- Adds a per-pixel uncertainty map, from SDE-variance sampling or an exact log-likelihood, so a radiologist can see where the model may have invented detail.
- Measured against a residual U-Net baseline on the Mayo Clinic low-dose CT data with SSIM and FID.
CodeSentry
AI code review for AI-generated code, shipped as a GitHub App. It pairs Semgrep (16 custom rules for common AI mistakes) with Claude behavioral analysis to catch bugs a static linter misses, then posts a risk-scored PR comment. LLM flags only survive if Semgrep backs them up, which kills most false positives.
Reproducible RL Pipeline
A template for deep RL experiments that actually reproduce: deterministic seeding, Hydra configs, version locking, experiment logging, CI, and Docker. PPO agents on CartPole, LunarLander, and Reacher, with mean ± std across seeds. Two runs of the same config give bit-for-bit identical results.
Nano Defect Detector
A lightweight PyTorch pipeline for spotting nanoscale defects in SEM and TEM images. It returns OK/NG classifications and pixel-level defect heatmaps in under 10 ms per 512×512 tile on a laptop, served through a FastAPI and Gradio viewer with an ONNX runtime path.
Personalized Writing-Style Tool
A multi-agent system that learns a person's writing style from samples and generates content that matches it with the Claude API, then runs a refinement loop against GPTZero feedback to keep the writing natural.
Writing
Calibration Precedes Evaluation
How we calibrate real-world robot evaluations with an adaptive difficulty staircase to learn more, spend less, and iterate faster.
Experience
From production computer vision to startup research and reinforcement learning for robots.
Summer 2026 – present
Software Engineering Intern
Parametric · San Francisco, CA · YC F25
Building and deploying reinforcement learning models for robotics, across the full ML lifecycle from training to production.
Winter 2026 – present
Research Assistant
Fan Lab, Stanford University · Stanford, CA
Contributing to MetaChat 2.0, a multi-agent framework for autonomous nanophotonic device design, and building the LLM evaluation set that measures it.
Summer 2025
Engineering Intern
Platform Science · San Diego, CA
Built a proof-of-concept model that computes Time-to-Collision from live vehicle video at 80% accuracy, using TensorFlow and OpenCV, with model training on AWS SageMaker and data in Snowflake.
Summer 2023
Product Manager Intern
Treeline Interactive · San Diego, CA
Worked across product development, market analysis, and roadmapping, and helped launch Five Iron Golf's online booking software.
About
In my free time, I love working out, golfing, seeing the world, and meeting new people. I'm also half Brazilian, so I try to go visit my family there as much as I can.
Education
Stanford University
B.S. Mathematics · Minor in Computer Science
2024 – 2028 · GPA 3.80 · Xfund Ethics Fellow · Stanford AI Club · Blythe Fund · Startups To Join
- CS 224R Deep Reinforcement Learning
- CS 231N Deep Learning for Computer Vision
- CS 109 Probability for Computer Scientists
- CS 107 Computer Organization & Systems
- Math 104 Applied Matrix Theory
- Math 53 ODEs, Linear Algebra & Fourier Methods
Skills
- Languages
- Python · Java · C++ · C · MATLAB · SQL
- ML & RL
- PyTorch · JAX · TensorFlow · Stable-Baselines3 · GRPO / PPO · Hugging Face · OpenCV
- Infrastructure
- Modal · AWS SageMaker · Docker · Fly.io · Snowflake · Hydra · W&B
- Product
- React · FastAPI · Postgres · Claude API
- Spoken
- English · Portuguese