SINCE 2025
Chan Zuckerberg Initiative
Staff Engineer AI/ML
- GPU-native computer vision
- Physics-grade ML evaluation
- Kernel-level inference systems
- Physics-to-production ML
ML Systems Engineer · CZI · Ex-NVIDIA
Miles Price
San Francisco Bay Area
The physicist who runs machine learning like an experiment — every claim dated.
Miles Price is a machine-learning systems engineer at the Chan Zuckerberg Initiative, where he designed and built the post-training and agent platform behind its education products end-to-end, as the sole engineer. At NVIDIA he created the 2,700-star CV-CUDA library and led its 20-engineer international team, and contributed to TensorRT's core. He came up as a computational physicist, mining petabytes of particle-accelerator data on supercomputers — and he flies as an FAA-certified private pilot.
- 2,700CV-CUDA GitHub stars
- ~20engineers led on CV-CUDA
- 40%faster Bi3D via custom CUDA operators
- 30%NGC deployment-efficiency gain
HAPPENING NOW
STANDING TODAY
HOW THEY THINK4 BELIEFS
Four beliefs carried intact from particle physics to LLM post-training.
THE IMPACT
From accelerator data mining to LLM post-training — pull one out.
EVERYTHING ELSE · 10 WORKS
THE TIMELINE
Five domains, one method: performance-first, evidence-obsessed, owned end-to-end.
The whole record, newest first – tap any moment and Aster tells the story.
2025–PRESENT
Building an LLM, model and harness, as the only engineer
At the Chan Zuckerberg Initiative since March 2025, Miles owns the ML systems behind foundation-model-driven education products. As the sole engineer he designed the full post-training stack — SFT with LoRA on expert-reviewed data, then GRPO refinement against rubric judges — plus a self-built async GRPO trainer, a knowledge-graph tool environment, and a measurement-first eval discipline with dated verification logs.
“Staff AI Engineer @ CZI, Ex-NVIDIA | LLM Post-Training (SFT, GRPO) · AI Agents · Evals · GPU Systems”
2020–2025
The library, the team, and the kernels under it
For roughly four and a half years Miles architected GPU inference and serving at NVIDIA. He created CV-CUDA from the ground up — an open-source C++/CUDA/Python library, 2,700 stars — and led its ~20-engineer international team, extended it into 3D, quantized Bi3D from FP32 to INT8 for a ~3× memory cut, contributed to TensorRT's core, and served as PIC for software product security.
“Machine Learning, Computer Vision @NVIDIA”
2018–2020
Vision that had to work on the factory floor
Between 2018 and 2020 Miles moved from research into deployment. At SoftWear Automation he built the computer-vision and sensor-fusion brain of the SewBot sewing robot, including its optomechanical imaging setup; at Stanley Black & Decker he shipped real-time audio-event, CCTV human-detection, and 2D/3D inspection systems across AWS, Google Cloud, and Azure, leading small teams on several.
2013–2018
Learned models inside safety-critical flight software
For five years at the Georgia Tech Research Institute, recognized as an SME in Operational Flight Programs, Miles built computer-vision simulations of embedded defense sensors — a full AAR-57 OFP simulation, an image-tracking SDK, a Grid-Engine simulation manager — and began replacing hand-coded detection logic with learned models: an SVM discriminator in the AAR-47, genetic-algorithm tuning that cut AAR-44 false positives.
2007–2013
Petabytes, supercomputers, and a physicist's evidence bar
Miles began in computational physics. As an undergraduate from 2009 he mined petabytes of particle-accelerator data on leadership-scale supercomputers, led the longitudinal analysis of the Θ+ particle at Jefferson Lab, and defended his work before the ~100-PhD CLAS Collaboration. His 2013 MS thesis studied electronic transport in graphene-nanoribbon heterojunctions. Performance engineering and dated evidence — the whole later career — start here.