About

Research, systems, and selected work.

Education

2023.09 - 2027.06
Bachelor, Automation
Xi'an Jiaotong University
Qian Xuesen College; GPA 3.7/4.3; class rank 2.
Recipient of the Yuejie Program Scholarship (35 selected from 3,500 students).
Coursework: Modern Control Theory, Multi-Agent Game Theory and Reinforcement Learning, Digital Image Processing, Linear Algebra.
2025.07 - 2025.09
Summer Visiting Program
University of California, Berkeley
Summer Session visiting student.
Completed AI & Scientific Engineering Applications and Academic English with A+ grades.
Built the Bi-Talk simultaneous-interpretation agent with international teammates and received an outstanding project distinction.

Internship Experience

Dec 2025 - May 2026
Research Assistant
A*STAR, Singapore
Researched egocentric video compilation with A*STAR/NUS collaborators, translating HD-EPIC annotations into executable symbolic environments for embodied planning and evaluation.
Jun 2026 - Sep 2026
AI Automation Specialist
Apple
Built and deployed an O2O budget-allocation agent using forecasting, contextual bandits, strategy adjustment, and A/B testing. Packaged data analysis, budget diagnosis, and allocation decisions as DeepSeek Harness plugins; the agent is used in the department’s daily operations.
Apr 2026 - Aug 2026
Research Assistant, GoTim Project Group
Peking University
Developed the interaction agent and vision system for a tutoring robot and handled system integration and perception debugging.

I am Qinchuan Cheng, an undergraduate in Automation at Xi’an Jiaotong University’s Qian Xuesen College. My research centers on Egocentric Compilation, LLM post-training and agents, multimodal foundation models, and AI for Science.

I study how agents can learn from real-world observations, compile first-person video into executable and verifiable environments, plan over long horizons, and recover from mistakes. My current work connects symbolic worlds, belief-state planning, branching experience, planner post-training, and agent-driven data synthesis.

Current Work

  • Egocentric Compilation — A research line spanning A*STAR/NUS collaboration, Ego2World, EgoVine, and EgoHarness: compiling egocentric video into executable worlds, branching experience, and training data for embodied agents.
  • LAW MASTER — Leading legal-agent development around evidence coverage, retrieval control, task decomposition, and multi-step reasoning.

Internships

Apple · AI Automation Specialist
Jun. 2026 – Sep. 2026
Built and deployed an O2O budget-allocation agent using forecasting, contextual bandits, strategy adjustment, and A/B testing. Packaged data analysis, budget diagnosis, and allocation decisions as DeepSeek Harness plugins; the agent is used in the department’s daily operations.

Peking University · Research Assistant, GoTim Project Group
Apr. 2026 – Aug. 2026
Developed the interaction agent and vision system for a tutoring robot, including system integration, gesture recognition, depth perception, relocalization, and projection interaction.

A*STAR, Singapore · Research Assistant
Dec. 2025 – May 2026
Collaborated with A*STAR and NUS researchers on the Egocentric Compilation line, compiling HD-EPIC annotations into executable environments for embodied planning and evaluation.

Snapshot

  • Xi’an Jiaotong University — B.Eng. in Automation, 2023–2027; GPA 3.7/4.3; class rank 2
  • UC Berkeley — Summer exchange program, 2025; A+ in AI & Scientific Engineering Applications and Academic English
  • Yuejie Program Scholarship — selected 35 out of 3,500 students
  • TOEFL — 100

Research Focus

  • Egocentric Compilation — Executable worlds, world graphs, belief-state planning, branching experience, and egocentric data synthesis.
  • Multimodal Foundation Models — Visual grounding, structured representations, and learning from real-world video.
  • LLM Post-Training & Agents — Reinforcement learning, retrieval control, tool use, and recovery in multi-step tasks.