CTO & Co-founder · Similate · a16z Speedrun SR7
Yubo Cao
I build swarms of LLM agents at Similate — and research how to evaluate them and keep them safe.
Caltech CS '29 · Student Researcher, Anima AI + Science Lab · prev. Lockheed Martin AI/ML
- 1000s
- concurrent LLM agents per simulation
- 10+
- studios running on Similate
- NeurIPS '26
- first-author paper under review
- ~10⁶×
- faster inference than Monte Carlo
Now
What I'm building right now
Similate — CTO & Co-founder
Thousands of synthetic users, one simulation
Similate (a16z Speedrun SR7) simulates real user populations with swarms of LLM agents to automate A/B testing — thousands of agents running concurrently per simulation, in production for 10+ design and brand-naming studios.
Research — LLM Safety & Evaluation
How do agent swarms fail?
Running thousands of production LLM agents is a front-row seat to how they break. My current research asks how to evaluate and align agent systems: benchmarks for synthetic-user fidelity, and fine-tuning studies where training loss improves while true ranking ability collapses — evidence that you must evaluate the capability you care about, not its proxy.
First-author paper — NeurIPS 2026 (under review)
Learning to outrun Monte Carlo
PTNO trains neural-operator surrogates for particle transport on cheap, noisy Monte Carlo labels — and ends up more accurate than its own training data, with inference up to ~10⁶× faster than converged Monte Carlo. Trained on 1M+ simulated configurations.
Side project — Live trading stack
Sub-second from social signal to order
Jiedusuo is a signal-relay and auto-trading platform I built and operate: an Axum Rust server on Postgres, a WebSocket feed, and a headless Python execution daemon, serving thousands of users at ~200k requests/day with ~0.5 s p50 end-to-end entry latency. The backtest lab behind it pre-registers every experiment, applies hard gates before a single maximand, and scores per-author edge with a recency-weighted EWMA with shrinkage before any strategy goes live.
Education
Caltech
Class of 2029, Computer Science. GPA 4.2/4.0. Learning Systems (A+), Decidability & Tractability (A), Experimental Robotics (A).
Georgia Institute of Technology
Dual Enrolled. GPA: 4.0. CS 1331 (A): Intro to Object-Oriented Programming; MA 1554 (A): Linear Algebra; MA 2551 (A): Multivariable Calculus; MA 3012 (A): Applied Combinatorics; and MA 2552 (A): Differential Equations.
Gwinnett School of Math, Sci, and Tech
GPA: 4.4, NGA: 101.960, Salutatorian. AP Computer Science A (5/5), AP Statistics (5/5), AP Computer Science Principles (5/5), AP Calculus BC (5/5), AP Physics C and E&M (5/5)
Experience
- Building a multi-agent synthetic-user simulation platform for automated A/B testing, running thousands of LLM agents concurrently per simulation.
- In production for 10+ clients across design and brand-naming studios; evaluation benchmarks curated from SimBench, with controlled experiments run before any customer-facing metric ships.
- First-author work (PTNO, under review at NeurIPS 2026): trained neural-operator surrogates for Monte Carlo particle transport from noisy, low-sample MC labels across 1M+ configurations, with inference up to ~10⁶× faster than converged Monte Carlo.
- Diagnosed the Jensen bias that appears when noisy labels are pushed through a log transform, and designed a positivity-preserving operator head and loss that predict fields spanning 9–11 orders of magnitude from noisy labels alone.
- Built the data-generation and training pipeline (Slurm, CUDA/NVIDIA Warp) behind the paper's 1M+ simulated configurations.
- Engineered "Security Advisor," a GenerativeAI assistant that won 1st Place in enterprise-wide Intern AI challenge. Projected to save ~28,000 engineering hours (approximately $2M+/yr) annually by consolidating security protocols into a unified, accessible portal.
- Built GenGNC, a LangChain multi-agent system that turns a high-level concept of operations into requirements and Simulink simulation prototypes, projected to cut prototyping time by 90%.
- Built emailing system for IT compliance tasks, Q&A system for controlled unclassified information handling, and AI agent for compliance audit, saving 80+ hrs/person/yr (approximately 400+ hrs/yr).
- Created application to fetch proposals to change regulations (dockets), saving 32+ hrs/docket (approximately 768+ hrs/yr).
- Migrated legacy Domino system to AI Factory, saving $2M+ in subscription costs and introducing MLOps practices in alignment with 1LMX.
- Rearchitected corporate-wide website from Angular to WordPress, saving 120 hrs/yr in maintenance costs.
- Fine-tuned BERT & used bipartite matching to match candidates to roles in internal leadership development program, saving 56 hrs/yr.
- Developed retrieval-augmented discrepancy reports system using StreamLit and AI factory for Solumina, automating discrepancy reports.
Open Source Contributor
Google Summer of Code (GSoC) · Jun 2025—now
- Selected as a contributor (top 1,300 globally) for JabRef (4k+ stars) with a $3,000 stipend.
- Enhanced onboarding experience, built declarative API for UI creation (similar to Tipkits), and event-driven rendering system in JavaFX.
Medlytics Student
MIT Beaver Works Summer Institute · Jun—Aug 2024
- Developed medical image visual question and answering system through CLIP and Word2Vec (ImageCLEF2023), achieving 80%+ accuracy and surpass the State-of-the-Art accuracy by 5%.
Founder & President
HackGwinnett · Oct 2022—Oct 2024
- Organized largest high school hackathons in Georgia three times ('23, '24, '25), with 200+ attendees, 50+ schools, and 10+ counties each.
- Hosted CS days for K-5 (200+ attendees) and raised $2,000+ partnering with State Farm & Amazon.
Independent Researcher
Gwinnett School of Math, Sci, and Tech · Sep 2022—May 2025
- Developed a novel locality-aware attention model for gene annotation, achieved near perfect accuracy, and presented at GA Junior Science & Humanities Symposium 2025.
- Developed gene therapy for chronic myeloid leukemia using CRISPR-Cas9 and chitosan polymers. Presented at American Society of Human Genetics 2024 conference (Board 8045F) as only high school project globally.
Skills
Languages
AI/ML
Numerical & HPC
Web
Cybersecurity
Infra & Tools
Awards & Honors
CalHacks 12.0 (UC Berkeley): Most Creative Hack, 1st of 695 teams ('25)
Carnegie Mellon University picoCTF: Top 8 of 1,498 US teams ('23); top 5% in US ('24)
LM CodeQuest: Top 3 in nation ('23) & 1st in state ('25); LM CyberQuest: 1st in state ('24, '25)
Air & Space Forces Association CyberPatriot: 4th in nation & 1st in state ('23); top 1% in nation ('24)
Riley's Way Call For Kindness: Fellow (top 50 globally), $5K grant for JobNest, a job platform for immigrants
Georgia Governor's Honors Program: Finalist for Software Engineering (top 13 statewide)
Coolidge Senator: Top 100 out of over 4,200 global applicants with $1,000 Scholarship ('24)
USACO: Silver division