Yufeng He · 何宇峰

Building reliableagent systemsfrom post-training to inference.

I worked on coding agents and multi-agent research as an intern at Moonshot AI (Kimi). I maintain Mooncake and contribute to vLLM and PyTorch, with a focus on the state, execution, and evaluation problems that make agents hard to trust.

Agent harnessesCoding-agent post-trainingMulti-agent systemsLLM systems
View résumé ↗
Emailheyufeng2003@qq.com
GitHubgithub.com/he-yufengLinkedInlinkedin.com/in/yufenghe
M.Sc. Computer Science · The University of Hong Kong
Yufeng He
Agent systemsOpen sourceHong Kong
540+upstream contributions
200+agent-framework & eval PRs
90+model-systems contributions
1.8K+CoreCoder stars

02 / Open source

Open source,
at systems depth.

540+ upstream contributions across 67 projects in the Agent and LLM ecosystem, from tool-call lifecycles to CUDA kernels and KV-cache transfers.

Maintained & contributed upstream

Contributions across
the Agent systems stack.

540+merged or landed PRs
across 67 projects
01

Frameworks & evals

200+

Merged upstream PRs

Qwen Code ↗56 PRsTool-call state & cancellation
deer-flow ↗24 PRsConcurrency & checkpoints
Microsoft Agent Framework ↗35 PRsAgent-loop boundaries
OpenClaw ↗8 PRsSession safety & permissions
02

Model systems

90+

Code, tests & tooling · Mooncake OSS Maintainer

Mooncake ↗56 PRs · MaintainerKV-cache correctness & transfer
vLLM ↗15 PRsAPI state & CUDA Graphs
PyTorch ↗16 landed PRsCUDA kernels & Inductor correctness
Triton ↗8 PRsGPU compiler semantics
03

Agent applications

100+

Merged upstream PRs

Vibe-Trading ↗64 PRs · MaintainerPortfolio analytics & trading gates
AstrBot ↗37 PRsProvider & scheduler reliability

Merged PRs and bot-landed changes; selected projects shown · Full contribution record ↗ ·

Independent open source01

Creator & Maintainer

CoreCoder ↗

Independently designed and built a compact Python coding-agent harness, covering tool execution, context management, sub-agents, and persistent sessions.

1.8K+GitHub stars

Python · Runnable agent harness

01 / Experience

Research depth.
Engineering range.

Across agent harnesses, coding-agent post-training, multi-agent systems, agentic finance, and LLM infrastructure.

Kimi

Moonshot AI

AI Agent Research Intern

Agent research that survives contact with engineering.

  • Researched coding-agent post-training, alignment, and harness design at Kimi Alpha Lab, from task construction to execution and trajectory analysis.
  • Independently built a long-horizon programming benchmark around data structures and performance optimization, with continuous feedback and model–harness comparisons; explored automated task generation, self-checking, and post-training for self-evolving agents.
  • Researched and implemented multi-agent and agent-swarm methods; independently ran experiments on task decomposition, candidate selection, and research–implementation–review workflows.
  • Core contributor to a quantitative-finance agent project, working on financial research automation, information synthesis, factor exploration, and experimental evaluation.
01

Baidu

AI Agent Engineer
AI Talent Program

Contributed to ads bidding that jointly modeled ROI and quality score, improving the team's business ROI target attainment from 81% to 101%.

02

Maimai

AI Agent Engineer
CEO Mentorship

Under CEO mentorship, independently built an agent pipeline for skill representation and quality evaluation, connecting unstructured profiles to automated career recommendations.

03

Kuaishou

Machine Learning Engineer
Transcoding scheduling

Optimized transcoding allocation across hundreds of thousands of CPU cores, jointly modeling content popularity, compute cost, and bandwidth savings.

05 / Research

Research with
testable claims.

My work connects coding-agent evaluation, collective learning in multi-agent systems, and AI-assisted mathematical discovery—building experiments and verifiable artifacts around each research question.

R/01

First author

TrajBias ↗

Built a coding-agent trajectory evaluation framework to investigate three classes of process-evaluation bias, with controlled comparisons across long-horizon task trajectories.

Submitted · ACL Rolling Review, October 2026

R/02

Project lead

AgentSociology

Built a multi-agent social-evolution testbed to explore group-level self-organization and continual evolution, with a focus on emergent swarm intelligence and collective learning.

Multi-agent systems · Group-level evolution

R/03

Project lead

AlphaFormal

Exploring a decentralized AI mathematical-discovery network where verifiable proofs establish trust between agents; connecting conjecture exploration, counterexample search, and proof reuse.

Three first-author manuscripts submitted to SCI-indexed mathematics journals · Lean checks of selected claims

R/04

Sole author

FMGAN ↗

Adapted adversarial refinement to time-series imputation model training, connecting adversarial learning with reconstruction of missing observations.

KDD 2026 MiLeTS

04 / Selected projects

Built beyond
the benchmark.

Two hands-on projects in health screening and AI music generation, combining model work with a usable application.

CTO · Jun 2025 — Jul 202601

Edge AI · Health screening

FormulaSperm

Built an edge-AI health-screening system from microscopy video, covering video-model training and on-device inference. Digital-twin simulation error was below 3%; the project won HKU DeepTech100.

Co-founder · Full-Stack Engineer
Oct 2021 — Jun 202202

Generative AI · Music systems

Qiyin Tech

Co-founded an AI music-generation project and developed its multi-agent collaborative engine and application. The project won National Gold at China's “Internet+” competition.

03 / Education

Sep 2025 — Nov 2026

The University of Hong Kong

M.Sc. in Computer Science

Sep 2021 — Jun 2025

Beijing University of Posts and Telecommunications

B.Eng. in Telecommunications Engineering

Selected honors

  • Champion2026 HKU DeepTech100 · HK$1M+ prize
  • Captain · 1st Place (Gold)BUPT · 2023 Shanghai Global AI Innovation Competition · LLM Track
  • 3× SilverACM ICPC Asia Regional Contests
  • National First PrizeNational Olympiad in Informatics in Provinces (NOIP)

03B / Skills & interests

Fluent across the stack.

Languages

Python / C++ (primary) · Java / Go

Agent systems

Agent harnesses · Coding-agent post-training · LLM systems

Practice

Claude Code / Codex power user · English (professional working proficiency)

Interest

National Top 10 · China Children's Go (Weiqi) Grand Prix