ResearchSep 2026
Loyalty Titration
Adding a second secret loyalty to expose a hidden one.
Co-author with Barry Shen
91.7% of answers followed the first hidden loyalty when a second one pushed against it
Early results: one model, one setting.
Leo Qin
I build AI-safety evals and web products, then test them against reality.
B.S. in Computer Science + Math, Washington University in St. Louis
Agent safety
HackathonJul 2026
Secret Loyalties
Installing, auditing and containing a hidden loyalty in a system prompt.
Team Elpis, with Barry Shen
60% of clean runs wrongly flagged as suspicious by a blind audit (15 of 25)
One weekend, one model and small samples: the report calls these preliminary results.
AI on real devices and robots
InternshipMay – Jul 2026
Violoop
Making a computer-use agent measurably better at its skills.
AI Systems Engineering Intern
1.5× average skill score, up to 2.5×, across 20+ skills
Scores come from a three-model AI judge panel.
Student teamSep 2026 – present
WashU Robotics Rover
GPS and IMU navigation software for a student rover.
Software Team Member (ROS 2)
Education and assistive tech
ProjectAug 2022 – Jun 2026
Personalized AI Teaching Copilot
A personalized AI assistant for teachers and students in Bijie, Guizhou.
Project Lead
543 textbooks in the library
A class-level comparison of two classes of about 50 seventh graders, not randomized. A formal analysis found no statistically significant overall difference in test scores, and the project does not claim a learning-effect improvement.
HackathonApr 2025
AI White Cane
Group Second Place at the Moonshot 48 Hour Hackathon Cup.
Team Lead
2nd place in the group, Moonshot 48 Hour Hackathon Cup
Maps and the language model need internet, and the board is too weak for real use yet.
Live website2024 – May 2026
BiphFlora
A bilingual campus plant catalogue.
Technical Lead
153 plant records with 970 photos
Live websiteAug 2023 – Jun 2026
StatsifyFinance
A bilingual finance learning platform.
Full-Stack Developer
100+ club members and public readers
Also learning
- A GPT from scratch
A character-level GPT in PyTorch: 10.8M parameters, 6 layers, 6 heads, trained on Tiny Shakespeare.
- ARENA exercises
CNNs, RNNs, RLHF and AI evaluation.
- BlueDot Impact
The Technical AI Safety course.
About
“My goal is to contribute to keeping AI agents safe.”
“I hope every student, wherever they live, can get an education built on their interests.”
- Education
Washington University in St. Louis
B.S. in Computer Science + Math- Experience
Violoop, an AI agent hardware startup in Shenzhen
AI Systems Engineering Intern, May – Jul 2026WashU Robotics rover team
Software Team Member (ROS 2), Sep 2026 – present- Focus
Agent safety
AI on real devices and robots
Education and assistive tech
- Outside work
- Badminton (WashU Badminton Club)
- Basketball
- Weightlifting
- Swimming
- Contemporary literature