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
117 verified merged author PRs across 28 upstream repositories, spanning robotics, machine learning, geometry and developer tools. Snapshot: 3 November 2026 · 22:08 CST (UTC+8).
Six projects, one row each, highlighting complementary work in program analysis, build-state management, numerical correctness and data handling. Each PR is verified as authored by Afloat16 and merged upstream. Rows follow their latest selected merge; dates use Asia/Shanghai. Technical details and scope.
| Latest selected merge | Upstream project · ★ | Representative engineering | PR / merged commit |
|---|---|---|---|
| 2026-11-03 | Craft Anything 6,553 ★ |
Persist rebuild progress across runs to stop unchanged attempts; compare WebAssembly tag signatures; recover minified esbuild dependencies through semantic bindings. | #557 · 48759fb#441 · fa4aae2#438 · 89bcd06 |
| 2026-11-02 | COLMAP 13,140 ★ |
Make benchmark translation-angle errors invariant to reconstruction scale through numerically robust vector normalization. | #4932 · 27ffad4 |
| 2026-11-02 | Sentence Transformers 19,251 ★ |
Preserve ranking-loss gradients with weighted log-sigmoid, stabilize FP16 scalar quantization, and keep padded LSTM outputs aligned with token features. | #4116 · abd266f#4113 · 847abf6#4301 · 26a0f49 |
| 2026-11-01 | DeepFace 23,642 ★ |
Align Milvus scores with DeepFace distance units, fetch complete galleries beyond the single-query limit, and preserve empty, singleton and multi-image batch structure. | #1634 · bf5ed73#1649 · 5696e33#1675 · 8e565c0 |
| 2026-10-27 | REA 425,500 ★ |
Complete partial ZIP reads and detect premature EOF; reject invalid UTF-8 in HTTP CDP discovery responses before JSON parsing. | #4414 · 82e4e03#4524 · 095a51c |
| 2026-10-22 | LeRobot 28,660 ★ |
Apply the existing V-JEPA 2 pairwise RoPE correction to VLA-JEPA's training-only world predictor. | #4835 · f9e9d73 |
18 upstream projects with verified top-30 positions. Craft Anything remains #2†, with 26 commits in its all-time contributor graph. Each rank carries its verification date; groups are sorted by the latest recorded stars.
| Project | Upstream ★ | Position · checked¹ | Merged PRs | Representative work |
|---|---|---|---|---|
| DeepFace | 23,642 | #7 2026-11-03 |
21 | Face-box clipping, crop boundaries, normalization and vector-search correctness. #1655 · #1658 |
| Sentence Transformers | 19,251 | #30 2026-11-03 |
5 | Ranking gradients, encoder masks and float16 quantization. #4127 |
| Denoising Diffusion | 10,724 | #9 2026-11-03 |
1 | Preserve external conditioning and self-conditioning in 1D diffusion. #373 |
| Craft Anything | 6,553 | #2† 2026-11-03 |
20 | Persist rebuild progress across runs to stop unchanged attempts; compare WebAssembly tag signatures; recover minified esbuild dependencies through semantic bindings. #557 · #441 · #438 |
| PyTorch Metric Learning | 6,343 | #9 2026-11-03 |
3 | Stable NT-Xent losses and gradients; independent HistogramLoss and PNP regression coverage. #758 · #759 |
| Atomic Agents | 6,316 | #28 2026-11-03 |
1 | Keep existing chat history intact when loading fails. #292 |
| Robotics Toolbox | 3,560 | #6 2026-11-03 |
11 | Kinematics, dynamics, inverse kinematics and URDF correctness. #801 |
| RSL-RL | 3,116 | #5 2026-11-03 |
3 | Resume RND schedules, preserve reward history and retain rollout remainders. #240 · #244 |
| PyPose | 1,620 | #15 2026-11-03 |
4 | Least-squares rigid alignment, robust-kernel stability and camera broadcasting. #437 |
| RL Games | 1,409 | #6 2026-11-03 |
4 | Retain the newest transitions in oversized replay batches. #384 |
| Geoopt | 1,103 | #3 2026-11-03 |
10 | Manifold constraints, gradient projection, optimizer state and stable geometry. #267 · #254 |
| skrl | 1,102 | #3 2026-11-03 |
3 | Gaussian policy math and cross-backend exploration consistency. #475 |
| pytransform3d | 707 | #15 2026-11-03 |
2 | Stable Lie-group Jacobians and screw-axis translation handling. #389 |
| RoMa | 647 | #2 2026-11-03 |
4 | Quaternion broadcasting, affine shapes and reproducible sampling. #33 |
| Rectified Flow | 490 | #3 2026-11-03 |
1 | Keep NanoFlow inversion in normalized flow coordinates. #50 |
| KinPy | 181 | #2 2026-11-03 |
3 | World-frame Jacobians with consistent axes and link offsets. #29 |
| Elements | 73 | #3 2026-11-03 |
2 | Preserve hidden filenames and correct local-path operations. #19 |
| qpmpc | 56 | #2 2026-11-03 |
3 | Preserve absent constraints, objective scaling and stable pendulum discretization. #15 · #17 |
| Project | Upstream ★ | Position · checked¹ | Merged PRs | Representative work |
|---|---|---|---|---|
| REA | 425,500 | #77 2026-11-03 |
3 | Deduplicate setup clients, complete ZIP reads and validate CDP response encoding. #4414 · #4524 |
| LeRobot | 28,660 | #228 2026-11-03 |
1 | Correct pairwise RoPE frequencies in VLA-JEPA's training-only predictor. #4835 |
| Haystack | 26,920 | #365 2026-11-03 |
1 | Keep pipeline snapshots reusable after a resumed run. #13076 |
| Recommenders | 21,975 | #49 2026-11-03 |
2 | Preserve custom columns and reject invalid top-k values. #2389 |
| COLMAP | 13,140 | #55 2026-11-03 |
2 | Scale-invariant translation-angle evaluation and benchmark path resolution. #4932 · #4931 |
| Kornia | 12,289 | #134 2026-11-03 |
3 | Finite geometry gradients and stable small-root quartics. #5517 |
| TileLang | 8,896 | #210 2026-11-03 |
1 | Honor the current CUDA device in profiling. #3357 |
| MLxtend | 5,190 | #95 2026-11-03 |
1 | Separate extracted emoticons from the preceding final word. #1207 |
| Trimesh | 3,715 | — | 1 | Correct nonuniform weighted Procrustes rotation while preserving equal-weight fits; merged into release/angles. #2618 |
| Geomstats | 1,528 | #84 2026-11-03 |
1 | Propagate product-metric parallel transport to each factor. #2131 |
¹ Unless marked †, positions follow GitHub’s commit-sorted REST contributor list, including bots and excluding anonymous contributors; ties retain API order. These 27 numerical REST positions retain their 3 November 21:43 snapshot. † Craft Anything uses the all-time Contributors graph for main, excluding merge commits, rechecked at 21:57: #2 / 26 commits; its separately observed REST value is #3 / 20. — means no numerical position was established. Six featured projects’ stars and Craft’s adopted work were refreshed in this update; other observations retain their recorded times. Full PR record, methodology and data.
TD-MPC2 (995 ★, contributor-list position #3): preserve target-Q history and evaluation mode during model conversion. The implementation from #98 was cherry-picked with Afloat16 authorship; the original PR was closed without merging. Together with the repositories above, this makes 29 upstream projects with accepted contributions.
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 117 merged author PRs. Details.
Craft Anything also adopted work from 32 original author PRs through six maintainer integrations: #645 · #648 · #649 · #652 · #653 · #656. All six default-branch commits retain my co-authorship; these adoptions are recorded separately from the 117 merged author PRs. Evidence.
Anti-REA Forge · MIT · 0.4.0 Alpha
Native release hardening with CMake-ready build integration, PE/ELF checks, guarded C++ memory and one-use request authorization. Explore runnable examples, cross-platform CI and research-guided design. Project story · Join the community.
ModelDock · v0.2.0
Local 3D/CAD preview with bounded 5+ GiB binary PLY/STL sampling, multi-model alignment and reference comparison. Validation record.
| Robotics | Large language models | 3D reconstruction |
|---|---|---|
| Bag Timeline Audit Read-only ROS 2 SQLite/MCAP timestamp diagnostics. |
LLM Handshake Bounded OpenAI-compatible endpoint probes and report comparisons. |
ReconTrail CPU photogrammetry, sparse geometry and offline reports. |
| TF Route Audit Exported TF and native URDF topology/time-window checks. |
SSE Toolcall Audit and Tooltrace Audit Saved stream and tool-execution diagnostics. |
Pose Convention Bridge and COLMAP Model Audit Explicit pose conversion and text-model preflight. |
Jev MCP · v0.5.0: a TypeScript MCP server for bounded probabilistic advice, routing and risk signals, with cross-client installers.
Research and engineering workflow: Research to Fix · Debate Direction · StewardCheck · Agent Action Notifier.
How I work: 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: 3 November 2026 · 21:43 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 | 3 | 12 | Python |
| Jev MCP | 1 | 49 | TypeScript |
| StewardCheck | 1 | 11 | Python |
| Debate Direction | 1 | 8 | Python |
| LLM Handshake | 1 | 6 | Python |
| ModelDock | 1 | 6 | TypeScript |
| Bag Timeline Audit | 1 | 5 | Python |
| ReconTrail | 1 | 5 | Python |
| Research to Fix | 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