I am He Dong, an M.S. student at the Beijing Institute of Technology, a former Agent Evaluation Infrastructure intern at ByteDance, and an external core contributor to Everything Claude Code. I build AI systems that can do useful engineering work while remaining observable, reviewable, and safe to operate.
My work sits at the intersection of LLM agents, AI for code, developer tooling, and distributed reinforcement learning. At ByteDance, I worked on AI-assisted engineering workflows that reduced an MR defect-tracing and code-review cycle from 30 minutes to 3 minutes.
observe → reproduce → locate → patch → verify
evidence at every boundary · humans retain the final decision
External core contributor · Everything Claude Code (ECC)
I contribute production changes across ECC's metrics pipeline, cross-process dashboard, hook runtime, installer lifecycle, and continuous-learning system. The work is review-heavy and regression-driven: bounded resource use, atomic state transitions, explicit failure semantics, cross-platform behavior, and focused test evidence are treated as part of the feature rather than follow-up cleanup.
| Engineering lane | Representative upstream work | Verified impact |
|---|---|---|
| Performance & observability | #3102 · session-cost snapshots #3064 · ECC2 output cursor |
Replaced repeated full-history reads with O(1) stable snapshots / O(delta) catch-up, measuring an approximately 3,099× steady-state speedup on a 40.5 MB fixture. Added a monotonic SQLite cursor for cross-process dashboard output. |
| Hook & install lifecycle | #2992 · Claude hook ownership #2987 · silent hook contracts |
Made manual hook installation ownership-aware and idempotent across install, upgrade, doctor, repair, and uninstall, backed by 177 focused tests. Preserved explicit hook output while eliminating accidental payload echo and bounding input across lifecycle entrypoints. |
| Failure containment & CI | #2971 · loss-safe archival #2194 · Unicode-safe CI |
Prevented semantic failures from authorizing observation archival, retained data across timeout/interruption paths, and unblocked the repository's Unicode-safety gate without losing localized rating information. |
Merged PRs · Upstream commits · Contributor graph
An evidence-first AgentTeam for repository maintenance. It moves an issue or failed CI run toward a verified pull request while preserving decisions, approvals, tool calls, and rollback points in a tamper-evident execution chain. Interactive demo ↗ · Verified PR |
A safe, example-driven codemod engine. Give it before-and-after snippets and it infers a structural transformation, scans a repository, previews the diff, and generates a rollback path before writing. Quick start ↗ |
A distributed Actor–Learner training system for MuJoCo with parallel rollout collection, PPO optimization, multi-GPU experiments, a Rust replay buffer, and real-time experiment observability. See the system ↗ |
A local-first CLI that makes Conda environment backup, restore, clone, offline packaging, cleanup, and diagnosis safer and more repeatable across Windows, macOS, and Linux. Project site ↗ · PyPI |
An AI presentation skill that produces animated, zero-dependency HTML decks and editable PowerPoint files through a deliberate theme system designed to avoid generic AI aesthetics. |
An agent-based system that turns product requirement documents into structured test-case suites, connecting product intent with repeatable quality workflows. |
| Principle | What it means in practice |
|---|---|
| Evidence before confidence | Reproduce the failure, capture the trace, and make every conclusion inspectable. |
| Safe autonomy | Give agents useful tools, explicit boundaries, human approval gates, and reversible actions. |
| Systems over demos | Build typed contracts, tests, observability, and deployment paths around the model. |
| Performance with a baseline | Measure against a frozen reference and promote changes only when regressions are understood. |
Languages
Agent systems & evaluation
Code intelligence & developer tooling
ML & distributed training
Backend & data
Reliability & delivery
I am currently exploring reliable agent infrastructure, repository-scale code transformation, evaluation systems, and scalable reinforcement-learning workflows. I am always interested in thoughtful collaborations around these areas.