Felipe Barbosa

Hey! I'm a Computer Science undergrad at Stanford University on the Artificial Intelligence Track, planning to graduate in June 2028.

I'm most interested in building systems that interact with their environment, especially through reinforcement learning, agents, and robotics. Within robotics, I spend a lot of time thinking about vision-language-action models and embodied reasoning. I'm also interested in post-training, inference, classical machine learning, and predicting complex systems.

The background is Langton's ant: a homage to some of the earliest steps in simulating complex systems.

Previously I led AI development at Onçafari, designed RAG pipelines and financial models at Atmos Capital, and researched molecular inhibitors at Stanford's Frydman Lab. I'm at Millennium Management as a Quantitative Research Intern.

Please reach out if any of these ideas sound interesting to you, and check out the other pages to learn more about me!

Millennium Management Jun 2026 – Present
Quantitative Research Intern
Atmos Capital Jun – Sep 2025
Summer Analyst
  • Designed production-grade RAG pipeline with Gemini embeddings, AlloyDB/pgvector, and LangGraph — boosted research efficiency 60%
  • Built real-time dashboard aggregating 80M+ nationwide insurance records to model market share and pricing trends
  • Conducted statistical pricing analysis using web scraping (BeautifulSoup, Selenium) and Pandas to identify price discrepancies across e-commerce platforms
Gemini, LangGraph, AlloyDB, pgvector, Google Cloud
Onçafari Jun 2022 – Sep 2024
AI Application Team Leader
  • Developed AI wildlife recognition processing 500+ hrs/mo of video, reducing manual review 95% for Latin America's largest conservation NGO
  • Secured Google Cloud support and led joint sprint with Google engineers
  • Curated largest Brazilian fauna dataset (1M+ annotated frames), eliminating ~6,000 hrs/yr of manual classification
Python, Google Cloud, CNN, Computer Vision, TensorFlow
Frydman Lab · Stanford Jan 2025 – Jan 2026
Research Intern
  • Computational modeling and virtual screening to design TRiC chaperonin inhibitors, followed by experimental validation
São Paulo State University Jan 2022 – May 2024
Research Intern
π-Drive, Real-Time Driving VLA
Awards: Honorable Mention in CS 224R (Deep Reinforcement Learning)
CS 224R final project adapting π₀.₅, a compact flow-matching VLA trained for robot manipulation, into a single-camera autonomous driving policy. Behavior cloning on NVIDIA PhysicalAI-AV teaches 6.4s acceleration-curvature trajectories; Flow-GRPO post-training improves mean ADE by ~13% and beats the best driving VLA, Alpamayo-R1 (3.58 m vs 4.23 m ADE), while running at control rate on a Jetson AGX Thor, ~3× smaller and ~20× faster. We're exploring applying it to our Stanford golf cart, Caddy.
Paper · Poster · Code
π₀.₅, Flow-GRPO, Behavior Cloning, PhysicalAI-AV, VLA, Modal
Caddy — Autonomous Golf Cart
Built and deployed a self-driving golf cart at Stanford, integrating perception, trajectory planning, and control for autonomous navigation in real campus settings. Demonstrated the system to Sam Altman, Jensen Huang, and Andrej Karpathy.
Mask2Former, Depth Anything V3, Motion Planning, Modal GPU Inference
PolySolver
Monte Carlo Tree Search agent enhanced with action filtering, heuristic rollouts, and a domain-aware value function for The Battle of Polytopia. Achieves 100% win rate against all public baselines. A Deep Q-Learning agent trained on PolySolver trajectories recovers 60% win rate with no search at inference time.
MCTS, Deep Q-Learning, MDP
911 Emergency Call Analysis
Real-time system that helps operators classify and prioritize emergency calls using AI. Analyzes call transcripts via multinomial inference and TF-IDF to determine emergency type, priority level, and key information — with Twilio integration for voice and SMS.
Multinomial Inference, TF-IDF, NLP, WebSocket