AI Researcher · Systems Builder
I build coding agents, large-scale multi-agent systems, and the infrastructure that turns LLM research into dependable software.
Shenzhen · China
00 · About
My education has unfolded alongside the rise of modern LLMs. I watched the field take off, and I expect to finish my master’s as it settles into a more mature, increasingly saturated phase.
I spent years following Microsoft’s XiaoIce (Xiaobing). It was the most capable conversational AI I had used at the time.
When OpenAI first opened GPT-3 to outside users, I applied and made it into the early cohort. My first impression was blunt: XiaoIce still felt better at conversation; GPT-3 felt like an astonishingly powerful autocomplete.
I joined GPT-4’s private preview. About a month later, OpenAI disabled the account during a wider enforcement sweep, shortly before broader access arrived. The account disappeared; my curiosity did not.
I was already building RAG systems with ChatGLM when the pattern was still new. I have stayed with the model family as ChatGLM grew into today’s GLM lineup.
I still test new models, probe their limits, and work out where they are genuinely useful. My daily stack includes Claude Max 20x and Codex Pro 20x. I was among the first annual subscribers to GLM Coding Plan Max, and I also keep an annual MiniMax Coding Plan subscription and a monthly Kimi membership.
OpenAI GPT · ChatGPT · Claude · Gemini · Llama · Mistral · Gemma · Grok · Cohere Command · Qwen · DeepSeek · ERNIE (Wenxin) · ChatGLM · GLM · InternLM · Kimi · MiniMax · Doubao · Baichuan · Yuanbao · Tiangong · Step · Xingchen · Hunyuan · Bailing · MiniCPM
01 · Experience
My work spans coding-agent evaluation, billion-agent simulation, retrieval, and long-context model training.
I study how repository history can help agents solve executable coding tasks, using controlled memory interventions and test-backed evaluation.
I worked on two systems: a planetary-scale social simulator and a production DeepResearch pipeline for sourcing, writing, citation tracking, and reliable delivery.
I worked on Jina Reranker v2 and Reader-LM, focusing on training data, hard-negative mining, long-context training, and benchmark design.
I conduct my research at the Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences.
02 · Selected Projects
I start with the failure mode, then work outward through data, runtime, evaluation, and delivery.
A native C++23 harness for coding agents, with multi-provider protocols, guarded tools, durable sessions, context recovery, subagents, workflows, MCP/LSP, and plugins.
A frozen retail sandbox for teaching compact models to search, compare products, check hard constraints, and complete multi-step shopping tasks through real tool calls.
A source-grounded study system that brings together document assembly, hybrid retrieval, cited answers, frozen evaluation, review planning, and PDF exam generation.
An LLM-driven social simulator that combines full models, distilled surrogates, columnar state, and batched events to run at planetary scale.
A courtroom simulator where lawyer agents sharpen their legal reasoning through structured adversarial exchanges across 1,000 civil cases.
A production DeepResearch pipeline that finds sources, plans reports, writes chapters, tracks citations, resumes failed jobs, and delivers long-form PDFs.
A recoverable multi-GPU batch pipeline around MinerU’s Python API. Spawned workers isolate CUDA contexts; workers stay pinned to GPUs while files flow through a shared queue; failures are logged without stopping the batch.
Four merged upstream PRs across LLaMA-Factory, Unsloth, and SGLang: Adam-mini across six training paths, dtype recovery after checkpoints, correct DeepSeek-Coder termination, and a Chat Completions compatibility fix.
A runnable Thought–Action–Observation loop for PocketFlow’s official cookbook. Four nodes, shared state, and explicit routing make each decision, tool result, and stopping condition easy to follow and extend.
A 278M multilingual cross-encoder built to rank documents, function schemas, tables, and code for agentic RAG. I proposed the function-calling and code-search directions and led their benchmark adaptation.
A family of compact long-context models that turns noisy HTML into clean Markdown. I contributed to the data, training, and evaluation pipeline, including two-stage 32K-to-128K training and seven model-scale studies.
03 · Publications
Selected work on coding-agent memory, GUI grounding, multi-agent systems, and trustworthy evaluation.
04 · Toolkit
A working range across research design, model training, agent runtimes, and production delivery.
Coding agents · Multi-agent orchestration · Tool use · Memory · Workflow runtimes
LLMs · Reranking · Long-context training · SFT · DPO · Evaluation
C++ · Python · PyTorch · FastAPI · Docker · Multi-GPU pipelines
Benchmark design · Controlled experiments · Failure analysis · Reproducible evaluation