ML systems engineer focused on JAX, model inference, and on-device AI.
I work on practical inference systems across JAX, PyTorch, CoreML, ONNX, WebGPU, and Swift. My current focus is lightweight video segmentation/tracking, edge deployment, and long-sequence streaming models.
- jax-ml/jax #27292 — Fix wrong unpacking in the decode attention Pallas kernel
Merged. Fixed an output-structure mismatch in the decode attention Pallas kernel that could cause shape mismatches and incorrect assertions whenbatch_size > 1.
-
patrick-kidger/equinox #235 — Support buffer donation in
filter_jit
Merged. Added support for controlling JAX buffer donation throughdonate_default,donate_args,donate_kwargs, anddonate_fn. -
patrick-kidger/equinox #233 — Fix
tree_inference
Merged. Fixed repeated node collection in nested PyTrees, reducing unnecessary traversal work in deeply nested module structures. -
patrick-kidger/equinox #190 — Fix attention mask handling and add tests
Merged. Fixed mask axis handling in attention and added regression tests.
Building lightweight video object tracking and segmentation pipelines for mobile and browser deployment.
Areas involved:
- EdgeTAM / SAM2-style video tracking
- CoreML model splitting and quantization
- Swift native inference
- ONNX / WebGPU deployment
- Mask generation for interactive visual effects
- JAX / PyTorch / Equinox
- Pallas kernels and inference correctness
- CoreML / ONNX / WebGPU
- Swift / iOS native inference
- Long-sequence and streaming model systems
- Edge AI product prototyping
I prefer building small, verifiable systems: model export, numerical comparison, runtime integration, and deployment on constrained devices.
