Leo Qin

← All work

03 · Internship

Violoop

Making a computer-⁠use agent measurably better at its skills.

Role
AI Systems Engineering Intern
When
May – Jul 2026 · Shenzhen, China
skills agent runs judge panel 123 three models reward reward from AI judges
The improvement loop: skills, agent runs, a three-⁠model judge panel, then updated skills.

Built with

  • Reinforcement learning
  • Computer-⁠use agent
  • Red-⁠teaming

About the project

At Violoop, an AI agent hardware startup in Shenzhen, I built a reinforcement learning pipeline that optimizes the skills of its computer-⁠use agent on AI-⁠judge rewards. Across 20+ skills, scores rose 1.5× on average and up to 2.5×, as rated by a three-⁠model judge panel.

I led 3 interns recording and scoring about 150 agent runs in 15+ scenarios, from WhatsApp follow-⁠ups to contract review. I also deployed a Discord support bot on a 432-entry Q&A knowledge base, red-⁠teamed it with 397 cases covering prompt injection and secret leaks, and fixed the gaps.

Worth knowing. Scores come from a three-⁠model AI judge panel.

Company code is private, so there are no screenshots or code links.

Next project