Selected Work

I work on the training, optimization, and deployment of deep learning models, with a focus on robotic perception and efficient execution on real hardware. The projects below describe my contributions to the teams and models linked in each section.

No Bias Stereo

Google (Intrinsic) · Stereo inference on edge accelerators

No Bias Stereo is a stereo-reconstruction model based on a Vision Transformer. My work focused on making the model run efficiently on Qualcomm NPUs.

Result: reached 10.76 FPS on dual Qualcomm NPUs, corresponding to 3.75× baseline throughput and exceeding the 10-FPS target for the evaluated robotics workload.

My contribution:

  • Rewrote ONNX and attention graphs for accelerator execution.
  • Applied operator fusion and selective post-training quantization.
  • Profiled and validated execution and numerical behavior on the target hardware.

These measurements refer to my Qualcomm deployment work. The public model paper reports its own GPU benchmarks, which use a different setup.

Public model project · Public model paper

3PTv2

Google (Intrinsic) · GPU acceleration and portability for multi-view 6D pose estimation

3PTv2: Zero-Shot 3D-Object Perception Transformer for Robotics estimates object poses from multiple views. My work focused on reducing execution time and exploring portability across hardware platforms.

Result: reduced inference latency on NVIDIA A100 from 6.8 to 3.1 seconds per frame, a 54% reduction.

My contribution:

  • Optimized CUDA reductions and reused vision-transformer computation.
  • Batched rendering and reduced CPU–GPU overhead.
  • Demonstrated a separate Qualcomm CPU/NPU proof of concept without retraining.

The A100 latency measurements and the Qualcomm portability proof of concept are separate results.

Public model paper

Flexion Reflect v0

Flexion Robotics · Vision-language agents and robot autonomy

Flexion Reflect v0 connects visual and language-based reasoning with executable robot skills.

My contribution: built agent components for autonomous robotic operation and integrated sim-to-real workflows, contributing to a successful demonstration on a physical robot.

In separate work on humanoid locomotion, I improved the existing training pipeline to achieve 2× faster training and 5% higher task reward.

Public project and team demo

Research and earlier experience

My academic work studies optimizer dynamics, hyperparameter scaling laws, compressed and distributed learning, and differential privacy. See publications.

Earlier, I worked on scalable machine learning at Yahoo Research and applied machine learning in finance at UBS. My CV contains the full experience and technical background.