Fudan University undergraduate · AI / Agents / Game AI / Optimization
I like building AI systems that can be run, tested, observed, and evaluated, not just demonstrated in a notebook. My recent work focuses on coding agents, AI product engineering, game AI, reinforcement learning, and real-world combinatorial optimization.
Current interests: LLM Agents · AI Coding · Agent Harness · Game AI · Reinforcement Learning · AI Product Engineering
| Project | What I built | Evidence |
|---|---|---|
| Agent Replay | A recorder, replay UI, evaluator, FastAPI service, and native ReAct coding-agent harness for Codex / Claude Code / OpenCode / Qoder / ZCode-style workflows. Includes tools, grader, retry, context compression, skills, MCP, reviewer, and multi-agent support. | Public beta; live demo; 130 live-model runs across 13 variants × 10 tasks; on one model and 10 Python tasks (one run per variant/task), context compression kept 10/10 verified success while using about 47% of full-context total tokens; not a general model ranking. |
| Beat Greedy | A playable ToC algorithm game that turns 0/1 knapsack into a player-vs-greedy-boss experience. Built the product flow, interaction, scoring, visual system, audio, and deployment. | React 19 + TypeScript + Vite + Tailwind; 18 automated tests; deployed online. |
| RealRoute-Bench | Research preview of a real-city routing benchmark: problem definitions, data provenance, an independent checker, and traffic-data acquisition; formal baselines and leaderboard remain incomplete. | 12 city worlds · 2,400 fixed road access points · N20/50/100/200 nested scales · 48 static duration/distance pairs. |
| Gorge Chase DIY Agent | DIY game agent built on Tencent KaiwuDRL: 16 actions, 1,840-dimensional observations, CNN + MLP Actor-Critic, GAE, masked PPO, and potential-based shaping. | Public code and training notes document environment interaction, learner updates, and saved checkpoints; no comparative performance claim. |
Replay what a coding agent did — then verify whether it actually succeeded.
- Record and replay real coding-agent sessions.
- Inspect edits, commands, tests, regressions, and recoveries on an interactive timeline.
- Compare multiple runs of the same task.
- Run a native Python ReAct harness with six built-in tools, independent grading, retries, context management, Skills, MCP, Reviewer, and Multi-Agent.
- Expose recording/evaluation through a FastAPI service.
- Evaluate system choices with measured experiments instead of relying on model self-reports.
Live demo: https://lyz1116-found.github.io/agent-replay/
Local optimum ≠ global optimum — turned into a small game.
The player packs a bag while a greedy shopkeeper only looks at value/weight ratio. After submission, the game reveals the player's solution, the greedy baseline, and the true dynamic-programming optimum.
- Product idea → interaction → implementation → testing → deployment.
- Drag-and-drop gameplay, difficulty and timing, scoring, boss feedback, local persistence.
- Original SVG assets and WebAudio sound design.
Play it: https://beat-greedy-da-e676.vercel.app/
I have hands-on experience with game-agent development using PPO / Actor-Critic / GAE / reward shaping / action masking, including a KaiwuDRL-based real-time game environment. I also received a National Third Prize in the Tencent Kaiwu Reinforcement Learning Competition.
I am a dedicated player of Honor of Kings (王者荣耀) and 王者万象棋, and an invited tester / real user of TapTap GameBuddy, and I am especially interested in AI products that combine visual understanding, domain knowledge, and real-time interaction inside games.
- LLM / Agents: ReAct, Function Calling, tool use, context management, evaluator/grader design, agent observability, prompt/agent experiments.
- ML / DL: PyTorch, Transformer, reinforcement learning, computer vision, conditional generation.
- Optimization: TSP/VRP-style combinatorial optimization, Gurobi, Concorde, LKH, OSRM-based routing pipelines.
- Engineering: Python, FastAPI, REST APIs, React, TypeScript, Vite, Tailwind CSS, Git, Docker, Linux.
- National Third Prize, Tencent Kaiwu Reinforcement Learning Competition
- University Second-Class Scholarship
- Academic performance: top 20% of cohort
I prefer projects where an AI system has to survive contact with reality: real tools, real state, real failures, real verification, and a user who needs to understand what happened.
For internships, I am particularly interested in AI product engineering, coding agents, agent infrastructure/evaluation, game AI, and AI-native consumer products.
Computer Vision Colorization currently publishes the course specification for grayscale-to-color generation in Lab space; implementation and experimental results are not included in that public repository.
Featured work above describes my project contributions. The Gorge Chase agent builds on Tencent's KaiwuDRL framework; I do not claim authorship of that framework. Forks, upstream mirrors, learning resources, and templates elsewhere in this account are not presented as original portfolio projects.