I build tools that turn hard-to-debug failures into reproducible, inspectable evidence.
_{把难以定位的问题，变成更小的复现案例、可执行测试与可审查的证据。}

Selected work · Research · Upstream contributions · Engineering approach · All repositories

Three open-source tools, connected by one idea: make a failure concrete enough to reproduce, inspect, and fix. These are early-stage tools with documented boundaries, not claims of production maturity or general model accuracy.

Reduce a failing browser journey to a verified reproducer and a runnable regression test.

ReproCut runs deletion experiments in fresh Chromium contexts and keeps a shorter action sequence only when the same declared failure still reproduces. It exports a replayable journey, a Playwright regression test, an offline evidence report, and a repair brief for a developer or coding agent.

Recorded demo: 12 → 3 actions · 36 browser replays · 75% fewer steps. One deterministic fixture, not a cross-application benchmark. Verified reductions are single-deletion minimal under the observed replay conditions, not guaranteed globally shortest.

JavaScript Node.js Playwright Delta debugging

Source & quick start · Demo artifacts & releases · 中文介绍

02 / RenderWitness Visual regression review, backed by inspectable evidence. Capture browser screenshots, isolate changed regions, and review the result with side-by-side, blend, and diff views. Includes scenario suites, ignore regions, explicit CI gates, and HTML / JSON / Markdown / JUnit exports. Optional VLM analysis stays separate from deterministic pixel evidence. The illustrated demo is metric-only; it does not establish real-VLM accuracy. Alpha tool. Python Playwright VLM adapters CI Source & demo · CI guide · 中文

03 / DreamFuzz Find, shrink, and replay world-model failures. Search for action sequences that expose model-vs-reference rollout drift, minimize the counterexample, and export a self-contained HTML replay. Uses explicit behavioral properties, paired open-loop rollouts, reproducible seeds, and an adapter contract for custom targets. The core has no third-party runtime dependencies. Alpha diagnostic tool, not a model-training framework or a general evaluation leaderboard. Python Property-based testing World models Source · Live replay · 中文

I am interested in reliable multimodal systems, long-video understanding, and learning from limited supervision. My focus is on making model behavior testable, evidence traceable, and evaluation reproducible.

_{Public software is listed above. Research manuscripts, results, and code are linked only when cleared for public release.}

I also work on focused bug fixes, regression tests, documentation, and code review in existing projects. Original projects, submitted patches, and review contributions are listed separately.

_{Status checked on September 8, 2026: the six PRs above were open. The linked upstream threads are the source of truth for later changes.}

LeRobot: reviewed the intermediate-prediction contract for world-model policies and reported regression risks around evaluation flag propagation and non-image outputs entering the video path. Review on PR #3757.

Jason Hu
_{AI reliability · World models · Multimodal systems · Open source}