I'm a Senior Machine Learning Software Engineer focused on reinforcement learning, ML systems, and research infrastructure.

I enjoy turning research ideas into working systems: implementing algorithms from first principles, designing controlled experiments, and building reliable infrastructure for machine learning.

kairos — Learning under real-time compute constraints

Investigates reinforcement learning when the environment cannot wait for the learner. Implements bounded-compute PPO with stale policies and dropped experience, using controlled experiments to study the trade-off between additional optimization epochs and fresher training data.

orbscreen — Uncertainty-aware materials screening at scale

Deep-ensemble graph neural network surrogate for predicting Orb-v3 relaxed-state MOF stability from unrelaxed structures. Halves energy MAE relative to a descriptor baseline, achieves ~6–7× top-decile enrichment, and uses selective simulation to reduce screening costs by 5–6.6×. Built using Modal, Orb-v3, and MofasaDB.

minitorch — Deep learning framework from first principles

Implements automatic differentiation, multidimensional tensors, neural network operations, and computation backends using NumPy, parallel Numba CPU kernels, and CUDA. Explores the abstractions and performance trade-offs underlying modern deep learning frameworks.

AlphaEx — Distributed experiment orchestration

Python tooling for launching parameter sweeps and coordinating thousands of Slurm jobs across research compute clusters. Designed to make large-scale experimentation easier to configure, execute, and reproduce.

• proprio — Unsupervised, uncertainty-aware perception for a 7-DOF robotic arm. Classifies lidar readings as self, background, or anomalies without relying on explicit robot geometry or kinematics.
• gym-puddle — Gymnasium Puddle World environment with an example off-policy reinforcement learning agent implemented using TorchRL.
• STOMP paper reimplementation — JAX implementation exploring reward-respecting subtasks and state construction.

• medical-jiwer — Medical speech-recognition evaluation built on JiWER, including medical-term WER, entity F1, and occurrence-aware scoring for doses, negation, and other clinically important values.
• BetterNǎo — Hackathon prototype exploring EEG foundation-model embeddings for wellness monitoring, with a React Native interface for viewing signals and annotations.
• priori — Interactive benchmark comparing tabular foundation models (TabPFN v2 and Google TabFM) against tuned XGBoost and AutoGluon on churn and credit datasets.

• internals — Interactive, first-principles explanations of modern AI systems, including a visual walkthrough of speculative decoding.
• nabla — NumPy implementations of 15 optimization algorithms, from SGD to Muon, with animated optimization trajectories and matrix least-squares benchmarks.
• map-reduce-distributed-grpc — Distributed MapReduce implementation using gRPC, with a coordinator dispatching map and reduce tasks to worker processes.

• cloze — Self-hosted language-learning application that generates fill-in-the-blank exercises from personal sentences and schedules reviews using SM-2 spaced repetition. Built with Next.js and PostgreSQL.

• Review of the Gemini Robotics 1.5 Technical Report