Hi, I'm Jash.
ML Research Engineer at Nissan. I like building things that people actually need.
| Project | Description |
|---|---|
| vla-serve | Real-time VLA inference for robots. 13x faster. One command. |
| robohacks26 | Pharmacy robot that builds semantic maps from video walkthroughs for autonomous navigation and manipulation. |
| vggt-factor-refinement | Factor graph refinement of VGGT for long video pose estimation. 70% ATE improvement over naive stitching. |
| recon3d | One-command 3D reconstruction from video. No COLMAP needed. |
| gtsam-splatfactors | Gaussian Splatting meets Factor Graph SLAM. iSAM2 pose optimization with differentiable rendering factors. |
| lerobot-doctor | Diagnostic tool for LeRobot datasets. Integrated into the official HF dataset visualizer. |
27 PRs merged across 12 repos.
PyTorch | GTSAM | LeRobot | JAX | PyTorch RL | Rerun | RSL_RL | Transformers | Accelerate | LeRobot Visualizer
PyTorch — 1 merged
| PR | Description |
|---|---|
| #182414 | Fix _replication_pad crash under deterministic mode with tensor padding |
GTSAM — 6 merged
| PR | Description |
|---|---|
| #2456 | Fix enforceConsistency_ in BatchFixedLagSmoother |
| #2463 | Fix incorrect key in cheirality error message for PPP/PPPC factors |
| #2464 | Fix tMin_ reset when an iteration has zero duration |
| #2465 | Reject negative sigma values in noise model constructors |
| #2467 | Fix LinearContainerFactor::rekey() segfault without linearization point |
| #2471 | Rename Python reserved keywords in wrapper interface files |
HuggingFace LeRobot — 5 merged
| PR | Description |
|---|---|
| #2486 | Enable torch.compile for DiffusionPolicy inference |
| #3016 | Replace assertions with proper exceptions in video frame decoding |
| #3017 | Fix missing config file in pip installs causing AttributeError crash |
| #3063 | Ensure queue and process cleanup on abnormal exit |
| #3485 | Fix LoRA resume from Hub (PosixPath cast + double wrap prevention) |
JAX — 1 merged
| PR | Description |
|---|---|
| #37358 | Add stable two-pass variance algorithm to nn.standardize |
PyTorch RL — 3 merged
| PR | Description |
|---|---|
| #3530 | Fix missing raise, incorrect __torch_function__ return, and off-by-one in RayCollector |
| #3589 | Fix MultiOneHot.to_numpy() returning scalar instead of per-space array |
| #3590 | Add set_at_, set_, update_ in-place write methods to ReplayBuffer |
Rerun — 3 merged
| PR | Description |
|---|---|
| #12672 | Fix incorrect byte unit conversion values in render benchmark |
| #12673 | Reuse precomputed timelines dict in send_columns instead of re-iterating |
| #12674 | Fix wrong variable in error message for extra args in log() |
RSL_RL — 2 merged
| PR | Description |
|---|---|
| #180 | Fix missing raise, wrong RND broadcast index, and variable typo |
| #181 | Fix mutable default arguments and replace assertions with proper exceptions |
HuggingFace Transformers — 1 merged
| PR | Description |
|---|---|
| #44287 | Fix mutable default arguments and resource leaks |
HuggingFace Accelerate — 1 merged
| PR | Description |
|---|---|
| #3944 | Fix mutable default in Megatron init and IndexError on empty ModuleList |
LeRobot Dataset Visualizer — 1 merged
| PR | Description |
|---|---|
| #57 | Add Doctor tab integrating lerobot-doctor diagnostics into the visualizer |