Projects
Systems, experiments, and things I’ve helped build.
Kernels & compilers
- Robotics Batch-Invariant Ops — Reusable PyTorch and Triton operators for batch-invariant inference in robotics foundation models.
- Vx — Contributing to a heterogeneous-systems language that makes hardware topology and memory placement part of the type system, with recent work on core language types and borrow-aware slices.
- FlashRT — Extended the high-performance VLA inference engine to AMD MI300X/CDNA3, including architecture-specific BF16 and FP8 kernels for low-latency robot policies.
- Batch-invariant π₀ — Tracing numerical divergence through vision encoders, dispatch, and batched matrix multiplication. Read the write-up.
Robotics & inference
- Deploying robots in datacenters — Deploying robots for datacenter operations. Optimized distributed VLA training on an 8-GPU node for 30% higher throughput, and developed and profiled BF16/FP8 CUDA, Triton, and HIP implementations for π0/GR00T-class models across NVIDIA and AMD GPUs.
- Orb — Built and instrumented real-time perception, eye-tracking, and feedback-control systems for the Orb, using hardware-in-the-loop evaluation to test tracking stability on physical devices.
- Custom GPU-Palletizing RL Environments — Reduced RL training time by 2× with distributed, GPU-parallel environments for palletizing. Developed synthetic-data generation and transformer/RL pretraining workflows, and implemented CUDA/C++ geometry and physics heuristics for learned combinatorial optimization.
- Teleoperation Infrastructure — Built low-latency real-time video streaming and teleoperation infrastructure, and optimized on-device perception for NVIDIA Jetson.
Production ML
- NVision — A manufacturing computer vision system built for live production lines, focused on edge deployment, real-time streaming, and practical operational value.
- Robin — A supermarket product scraping and matching system for large-scale catalog retrieval, normalization, and embedding-based product similarity.
Learning systems
- Continual learning & online distillation — Studied SDPO-based continual learning and online distillation, focusing on training throughput and reliable post-training execution.
- Numerical reliability of robotics foundation models — Identified batch-dependent inference drift across 10+ robotics models and demonstrated changes in task outcomes. Developing reusable operators for numerically consistent robot inference.
- DreamZero: high-throughput WAM rollout engine — Designing a distributed, asynchronous rollout system for world-action-model RL, focusing on high-throughput sampling, actor/learner execution, and numerical parity between rollouts and training.
- Low- and mixed-precision adaptation of VLAs and WAMs — Investigating FP8/FP4 full-model and LoRA fine-tuning, measuring training efficiency, optimization stability, and downstream policy behavior.
- Reconstructive memory — A cognitive science framing of memory as reconstructive inference in learning systems.
- CORE — Full-path evaluation of LLM agents beyond their final state.
- Agricultural agents — Deploying and evaluating agents over full-season soybean farm operations.