My research focuses on robotics, large language models, and 3D reconstruction.
I build reproducible research tools, explore native software protection, and contribute focused upstream fixes.
Python · TypeScript · C/C++ · Reproducible experiments · Upstream engineering
231 verified merged author PRs across 33 upstream repositories, spanning robotics, machine learning, geometry and developer tools. Snapshot: 9 November 2026 · 20:33 CST (UTC+8).
Homepage contribution lists show projects with at least 1,000 stars. The totals above include all verified merged author PRs, regardless of repository stars.
Seven projects, one row each. Each row highlights one to three substantial recent merges from that project's contribution history, selected for their engineering impact and regression evidence. Projects follow their latest selected merge, with dates in Asia/Shanghai. Technical details and scope.
| Latest selected merge | Upstream project · ★ | Representative engineering | PR / merged commit |
|---|---|---|---|
| 2026-11-09 | Craft Anything 23,755 ★ |
Decode concatenated Zstandard frames under one output limit, identify EVM dispatchers that actually reject payments, and keep filesystem recovery inside declared image bounds. | #1340 · 25a8f17#1323 · 3f7b9a3#1246 · b9a3525 |
| 2026-11-09 | Claude-Mem 115,178 ★ |
Protect the selected restore point during backup rotation, preserve distinct files whose hash prefixes collide, and keep retried PostgreSQL writes idempotent. | #5532 · 5ce4b19#5536 · e15d466#5538 · 5047f2b |
| 2026-11-09 | COLMAP 13,183 ★ |
Restore Python 3 image collection with bounded retries and response cleanup, and make translation-angle evaluation invariant to reconstruction scale. | #4977 · 9f0736e#4932 · 27ffad4 |
| 2026-11-08 | DeepFace 23,662 ★ |
Preserve explicit false API options, close HTTP image responses through success and failure, and keep empty, singleton and multi-image batches structurally consistent. | #1684 · a8cee7b#1679 · eb29c2e#1675 · 8e565c0 |
| 2026-11-06 | REA 445,495 ★ |
Preserve UTF-8 at diagnostic cutoffs, resume complete NDJSON messages after interrupted iteration, and honor cancellation when bounded reads finish. | #8907 · 3eff806#8909 · a3621e5#8908 · b153ba8 |
| 2026-11-05 | Sentence Transformers 19,269 ★ |
Exclude padded tokens before CNN convolution and compute RRF retrieval metrics without inserting unretrieved positives into observed rankings. | #4323 · 7da542f#4114 · 6994c94 |
| 2026-10-22 | LeRobot 28,772 ★ |
Apply V-JEPA 2's pairwise RoPE correction to VLA-JEPA's training-only world predictor. | #4835 · f9e9d73 |
15 displayed upstream projects have verified top-30 positions in GitHub's current Contributors graphs. Positions were checked on 9 November 2026 using All / Commits / default branch / excluding merge commits. Groups are sorted by the latest recorded stars.
| Project | Upstream ★ | Position · checked¹ | Merged PRs | Representative work |
|---|---|---|---|---|
| Claude-Mem | 115,178 | #16 2026-11-09 |
15 | Keep a backup selected for a pending restore until the restore marker clears. #5532 |
| Craft Anything | 23,755 | #2 2026-11-09 |
86 | Decode concatenated Zstandard frames under one cumulative output limit. #1340 |
| DeepFace | 23,662 | #4 2026-11-09 |
25 | Clip detected face boxes before NumPy slicing and preserve false boolean options in HTTP form requests. #1655 · #1684 |
| Sentence Transformers | 19,269 | #28 2026-11-09 |
6 | Ranking gradients, FP16 quantization and padding-safe CNN/LSTM encoding. #4323 · #4116 |
| COLMAP | 13,183 | #28 2026-11-09 |
5 | Fix the IMC 2024 dataset layout for relative paths and resolve benchmark paths before changing subprocess directories. #4976 · #4931 |
| Denoising Diffusion | 10,724 | #9 2026-11-09 |
1 | Preserve external conditioning and self-conditioning in 1D diffusion. #373 |
| ThinkRail | 8,030 | #14 2026-11-09 |
3 | Identify SVG by its XML root, and keep UTF-16 surrogate pairs intact when truncating terminal output. #1184 · #1186 |
| PyTorch Metric Learning | 6,342 | #5 2026-11-09 |
8 | Accumulate class-weighted losses in FP32 before returning the input dtype; keep center-regularization losses in FP32 for half-precision inputs. #775 · #776 |
| Atomic Agents | 6,321 | #23 2026-11-09 |
1 | Keep existing chat history intact when loading fails. #292 |
| Robotics Toolbox | 3,570 | #7 2026-11-09 |
11 | Kinematics, dynamics, inverse kinematics and URDF correctness. #801 |
| RSL-RL | 3,124 | #4 2026-11-09 |
3 | Resume RND schedules, preserve reward history and retain rollout remainders. #240 · #244 |
| PyPose | 1,621 | #16 2026-11-09 |
4 | Least-squares rigid alignment, robust-kernel stability and camera broadcasting. #437 |
| RL Games | 1,411 | #7 2026-11-09 |
4 | Retain the newest transitions in oversized replay batches. #384 |
| skrl | 1,105 | #3 2026-11-09 |
5 | Reduce noncontiguous PyTorch gradients across workers and propagate JAX worker failures to the launcher. #489 · #496 |
| Geoopt | 1,104 | #2 2026-11-09 |
13 | Correct canonical Stiefel tangent projection, Cayley initial velocity and line-search derivatives; preserve stable manifold optimization. #280 · #267 |
| Project | Upstream ★ | Position · checked¹ | Merged PRs | Representative work |
|---|---|---|---|---|
| REA | 445,495 | #59 2026-11-09 |
6 | Complete partial ZIP reads and detect premature EOF; preserve UTF-8 code-point boundaries in byte-bounded diagnostic output. #4414 · #8907 |
| LeRobot | 28,772 | — 2026-11-09 |
1 | Correct pairwise RoPE frequencies in VLA-JEPA's training-only predictor. #4835 |
| Haystack | 26,951 | — 2026-11-09 |
1 | Keep pipeline snapshots reusable after a resumed run. #13076 |
| Recommenders | 21,984 | #46 2026-11-09 |
3 | Preserve custom columns in k-core filtering and stratified splitters; #2419 merged into staging. #2389 · #2419 |
| CARLA | 14,581 | #81 2026-11-09 |
3 | Correct the documented change of basis for 4×4 poses and distinguish the precision of packed depth measurements from grayscale images. #10188 · #10189 |
| Kornia | 12,361 | #85 2026-11-09 |
6 | Accumulate weighted FP16/BF16 meter updates in FP32; export BF16 images to NumPy without changing channel layout. #6580 · #6582 |
| TileLang | 8,933 | — 2026-11-09 |
1 | Honor the current CUDA device in profiling. #3357 |
| openMVG | 6,593 | #96 2026-11-09 |
1 | Use depth along each observed bearing for DLT triangulation, preserving valid rear-hemisphere observations from spherical cameras. #2445 |
| MLxtend | 5,192 | #60 2026-11-09 |
1 | Separate extracted emoticons from the preceding final word. #1207 |
| Trimesh | 3,719 | — 2026-11-09 |
1 | Correct nonuniform weighted Procrustes rotation while preserving equal-weight fits; merged into release/angles. #2618 |
| Manifold | 2,400 | #76 2026-11-09 |
1 | Keep static convex-hull construction independent of input execution contexts while preserving cached operand errors. #1905 |
| Geomstats | 1,529 | #83 2026-11-09 |
1 | Propagate product-metric parallel transport to each factor. #2131 |
¹ Positions are the explicit ranks displayed by GitHub's official Contributors graph, read on 9 November 2026. Current comparisons with the previous snapshot use the same graph settings. Earlier REST observations remain in the full record and are not directly comparable. — means Afloat16 was absent from the complete displayed top 100; it is not an exact rank or a claim of zero contribution. GitHub may cache graph calculations. Graph counts, default-branch author commits and merged PRs are separate metrics. Full record, methodology and data.
Robotics Toolbox also incorporated the circular particle-heading work from #735 into #883, and the trajectory work from #736 into #805, retaining co-authorship. These adoptions are recorded separately from the 231 merged author PRs. Details.
Craft Anything incorporated work from 101 of my original PRs in 11 maintainer integrations. The latest, #1217, combined seven changes to format readers and project launchers, retaining my co-authorship. An earlier squash, #995, omitted my trailer; the maintainer restored attribution separately in #1012. All 11 integrations and their sources.
Claude-Mem adopted the invalid-input regression cases from #5339 into #5336, retaining my co-authorship. My adopted contribution covers the rejection tests; the production numeric-string conversion was implemented by the other PR's author. This adoption is separate from my 15 directly merged PRs. Evidence.
Small tools for concrete engineering problems. Explore a project, try its examples, and star it if it fits your work.
|
Anti-REA Forge Separate private ELF symbols at build time and protect service operations with request-bound, one-use capabilities. Inspect paired Ghidra and IDA fixtures.
|
ModelDock Preview, align and compare meshes or point clouds locally. Explore sampled binary PLY/STL previews above 5 GiB and export comparison reports.
|
Jev MCP Bring TypeSafe Jev advice to bounded choices, routing and risk checks through four MCP tools, with installers for multiple coding clients.
|
|
Bag Timeline Audit Find timestamp gaps, reversals and duplicates in ROS 2 SQLite or MCAP recordings, then compare topics using an explicitly shared clock.
|
LLM Handshake Probe OpenAI-compatible endpoints for streaming, tool calls and structured output. Preview exact requests and compare reports after a gateway or model upgrade.
|
ReconTrail Check photo quality, reconstruct sparse 3D geometry on a CPU, and share a self-contained interactive report that opens offline.
|
|
TF Route Audit Check exported TF records and native URDF for cycles, disconnected frames and conflicting edges. Query routes against observed time windows.
|
SSE Toolcall Audit Reassemble parallel tool calls from saved SSE captures and check completion markers, JSON arguments and consistency with a saved final response.
|
Pose Convention Bridge Convert camera poses between OpenCV and OpenGL axes or world-to-camera and camera-to-world forms, with validated transforms and Nerfstudio export.
|
|
COLMAP Model Audit Audit COLMAP text models for camera-field errors, broken reciprocal tracks and disconnected view graphs, with optional image-path and model-overlap checks.
|
Tooltrace Audit Review saved execution traces for overlapping retries, duplicate attempts and unmatched results. Add the optional recorder around your own tool calls.
|
Research to Fix Research similar failures, map relevant solutions to your code and verify the repair. Save cases with source links and before-and-after evidence.
|
|
Debate Direction Let two agents propose, challenge and revise a plan. Keep the objections, remaining disagreements and next validation steps in a reviewable report.
|
StewardCheck Pin a task baseline, inspect edits against the agreed scope and keep test receipts tied to the files they actually checked.
|
Agent Action Notifier Turn explicit workflow events into queued email or desktop alerts, with pending-request tracking, retries and recovery after interrupted delivery.
|
I isolate the symptom, understand the boundary, write a regression, and make the result inspectable. The six domain utilities have 67 local tests and 36 passing OS/Python CI combinations at the recorded release. These are experimental diagnostics and conversion tools; each repository documents its limits.
All 15 projects: stars and default-branch commits
Snapshot: 9 November 2026 · 20:47 CST (UTC+8). Sorted by stars, then repository commit count. Commits count reachable default-branch history, including merge commits; they are not personal authored-commit counts. Sources and versions.
| Project | ★ | Default-branch commits | Main language |
|---|---|---|---|
| Anti-REA Forge | 6 | 29 | Python |
| Jev MCP | 1 | 49 | TypeScript |
| StewardCheck | 1 | 11 | Python |
| Debate Direction | 1 | 8 | Python |
| Research to Fix | 1 | 7 | Python |
| LLM Handshake | 1 | 6 | Python |
| ModelDock | 1 | 6 | TypeScript |
| Bag Timeline Audit | 1 | 5 | Python |
| ReconTrail | 1 | 5 | Python |
| SSE Toolcall Audit | 1 | 5 | Python |
| TF Route Audit | 1 | 5 | Python |
| Tooltrace Audit | 1 | 5 | Python |
| Agent Action Notifier | 0 | 6 | Python |
| COLMAP Model Audit | 0 | 4 | Python |
| Pose Convention Bridge | 0 | 4 | Python |
Project guide · Demand research · Project metrics · Contribution record