Hello, I’m He Dong.
I build agent infrastructure, evaluation systems, and developer tools that remain observable, reviewable, and safe to operate—from production engineering platforms to open-source experiments.
Engineer, researcher, builder
I’m an M.S. student at Beijing Institute of Technology, focusing on deep reinforcement learning and agentic RL. During my B.Eng. in Computer Science at Hebei University of Technology, I ranked 1st of 139.
Across two ByteDance internships, I worked on the infrastructure behind trustworthy AI delivery: defect tracing, automated code review, LLM-as-Judge pipelines, evaluation datasets, release gates, and the platform surfaces that make all of them usable.
Outside work, I turn the same ideas into open-source tools—from auditable repository agents and AST codemods to distributed reinforcement-learning systems.
From evaluation contracts to production systems
A timeline of building engineering infrastructure, evaluation pipelines, and research systems—always close to the real failure mode.
ByteDance · Data Platform · iDA
Led the 0→1 build of EvalHub, bringing performance benchmarking, skill evaluation, baselines, Meego releases, and Bits governance into one React + Fastify platform. Extended the shared evaluation pipeline with multi-turn skills, non-text artifacts, release gates, and 70-case / 72-turn regression baselines.
ByteDance · PDI · China Commerce & Ads
Built an Aime Workflow → PE → Skill system for MR defect tracing and automated code review. The production workflow reduced one review cycle from 20–30 minutes to 2.5–3 minutes, with 98% localization accuracy, 95%+ decision consistency, and under 3% false positives.
Institute of Artificial Intelligence · HEBUT
Worked on multi-objective route planning for battery-swapping heavy trucks, covering problem formulation, experimental validation, and paper collaboration. The work was accepted by NTCI 2025.
Open source as an engineering record
Select a project to inspect its premise, mechanisms, and public evidence.
Evidence-first repository maintenance
A six-agent system that moves issues and failed CI runs toward verified pull requests, while retaining an append-only evidence chain, human approval gates, and rollback points.
A coordination agent that interrupts less
A privacy-aware group coordination agent built with Strands. It combines a commitment ledger, earned autonomy, minimum-quorum routing, and a hard interrupt budget.
Structural change by example
A Rust CLI that infers AST-level transformations from before-and-after snippets, previews every match as a diff, and creates a rollback path before writing.
− fetchUser({ userId: id })
+ getProfile({ profileId: id })Safer environment operations
A cross-platform CLI for backup, restore, clone, offline packaging, cleanup, and diagnosis—wrapping native Conda with safer defaults and actionable errors.
$ conda-helper doctor
✓ conda installation
✓ writable cache
✓ environment metadata
All systems healthy.Distributed reinforcement learning
A Ray-based Actor–Learner system for MuJoCo with PPO, multi-GPU scaling experiments, a Rust replay buffer, and real-time training observability.
Reviewed in public, improved in the open
My open-source work focuses on repository-scale reliability, installation paths, data correctness, and secure integration behavior—each contribution linked to its public review trail.
A top-10 contributor to ECC, with work merged across performance-critical data paths, hook security, installation integrity, failure-safe learning, and cross-platform CI. Each case below links to its upstream review and verification record.
Each contribution starts from a concrete failure mode, changes a production path, and carries its own verification evidence.
Invited by Future AGI founder Nikhil Pareek to contribute to the project. The invitation came through private communication; the links alongside point to the resulting public engineering work.
Founder reference ↗Evidence beyond the repository
“Confidence is useful only when the evidence survives review.”
GPA 3.91 / 4.0 · ranked 1st of 139 in Computer Science.
Provincial recognition for academic performance and leadership, alongside Outstanding Student Leader and Outstanding Communist Youth League Member honors.
Team leader for the national undergraduate mathematical modeling competition.
A featured dataset with 3,000+ downloads, plus practical work across AI competitions.
Published a technical article on ByteDance’s internal ByteTech forum; it ranked No. 1 on the homepage weekly chart. The article is internal and has no public link.
Contact
The easiest way to reach me is by email. You can also find my public work on GitHub.