HD He Dong AI Systems / Agent Engineering
GitHub

M.S. Student at BIT · Former Agent Eval Infra Intern at ByteDance

I build AI systems that leave evidence behind.

I am He Dong, an engineer working across coding agents, evaluation infrastructure, and developer tooling. I turn ambiguous work into traceable systems that can show what changed, why it changed, and how it was verified.

Explore selected work github.com/wellkilo
EXECUTION TRACE
run_042
  1. 01
    Issuecapture the real failure
    observed
  2. 02
    Triageseparate signal from noise
    scoped
  3. 03
    Locatefollow the actual call chain
    traced
  4. 04
    Patchmake the smallest safe change
    changed
  5. 05
    Verifyprove the intended behavior
    passed
Evidence coverage
5 / 5
Scroll to inspect

Selected work / 精选项目

Systems, not demos.

Projects built around reliability, developer control, and inspectable outcomes.

01 / AgentTeam TypeScript

Flagship project

RepoPilot

Evidence-first autonomous repository maintenance.

An auditable AgentTeam that moves a GitHub issue or failed CI run toward an independently verified pull request, preserving decisions, tool calls, approval gates, rollback points, and validation evidence along the way.

IssueTriageLocatePatchVerify
  • AgentTeams
  • GitHub
  • Audit trail
  • PR-only safety
View repository
02 / Developer toolRust

codemod-pilot

Transform a codebase by example.

Learns structural transformations from before-and-after code, scans for matches, previews unified diffs, and applies changes with rollback support.

  • AST
  • Safe preview
  • Rollback
GitHub
03 / Open courseTypeScript

Learn Harness Engineering

The environment around the agent is part of the product.

A multilingual, project-based course on the state, verification, and control mechanisms that make long-running coding agents dependable.

  • 12 lectures
  • 6 projects
  • Agent harness
GitHub
04 / AI testingTypeScript

CaseAI

AI-assisted test design with human control.

A test-case generation and management system with document parsing, multi-model fallback, versioned edits, audit history, visual maps, and structured exports.

  • React
  • LLM routing
  • Audit log
GitHub

Engineering approach / 工作方式

A result is only useful when its path can be inspected.

01

Trace the real system

Start from the user-visible failure, then follow the actual contract and call chain before changing code.

02

Keep the change narrow

Prefer a small, reversible intervention over a broad rewrite. Preserve compatibility where the public surface matters.

03

Verify both sides

Test the target behavior and a negative control. The best evidence proves what changed and what did not.

Portrait of He Dong wellkilo

About / 关于

He Dong

M.S. student at Beijing Institute of Technology and former Agent Evaluation Infrastructure intern at ByteDance.

My work sits where AI agents meet software engineering: evaluation contracts, long-running harnesses, repository automation, and tools that make large changes safer. I care less about a system appearing autonomous than about it being understandable, recoverable, and useful to the engineer responsible for the outcome.

Current focus
Reliable coding agents and evaluation
Working across
TypeScript, Python, Go, Rust
Based in
China · UTC+8

Contact / 联系

Have a hard engineering problem?

Bring the context. I will bring the trace.

Start on GitHub