ENGINEERING / PUBLIC REPOSITORIES

I test ideas by building the system.

I came into engineering through machine learning, then kept moving down the stack. When a project depends on a database, GPU backend, or simulator, I want to understand what it is doing and where it can fail.

Public GitHub

SELECTED BUILDS

What I have been working on.

Rust · ROCm / HIP

candle-rocm

A ROCm/HIP backend for Hugging Face Candle, built below the model layer with device discovery, safe Rust wrappers, GPU kernels, and rocBLAS matrix multiplication.

  • Custom HIP runtime and FFI crates
  • Used as the AMD GPU backend in picochat

Rust · language-model training

picochat

A Rust framework for language-model training and inference. It brings tokenizer training, pretraining, fine-tuning, reinforcement learning, and interactive use into one workflow.

  • candle-rocm integrated as a pinned GPU backend
  • Training and inference designed to stay close to the system

MuJoCo · PPO / behavior cloning

G1 manipulation challenge

A pick-and-place submission for a Unitree G1 simulation. The work splits the task into environment design, a scripted teacher, behavior cloning, and PPO fine-tuning.

  • Task-agnostic training framework
  • Committed ONNX routine for headless evaluation

TypeScript · Bun · Cloudflare R2

GitForge

A multi-tenant Git hosting platform that supports Git Smart HTTP and Git LFS, with auditable controls around access, protected resources, and administrative activity.

  • Tenant-isolated object storage
  • Role-based access control and audit logging

C++23 · concurrent systems

ScalerDB

An in-memory database built to learn the parts that libraries normally hide: schema checks, primary-key indexing, persistence, read-write locking, and latency measurement.

  • std::variant value system
  • GoogleTest suites for data, persistence, and concurrency

R · model validation

Diabetes prediction

The model reached 1.00 AUC because the target was derived from fasting glucose while fasting glucose remained an input. That result was a leakage warning, not a model win.

  • 10,000-row public dataset
  • Limitation documented in the project README

R · group research

NFL career data project

A group analysis of quarterback statistics and career win percentage using Lasso and stepwise regression. The test results explained about 37–38% of the observed variation.

  • Career game logs aggregated by starter
  • Omitted team context stated as a limitation
GROUP PROJECT / CONTRIBUTOR