THE MAHOGANY WRITING BUREAU // OPEN-SOURCE & RESEARCH CODE
From creating Teras (multi-backend tabular deep learning across JAX, PyTorch, and TensorFlow) and contributing random-sampling APIs to official Keras 3, to porting Gemma 2from Flax Linen to NNX for staged distillation (codriver) and writing an autograd engine from scratch in pure C (micrograd_c).
BUREAU DRAWER I // OPEN-SOURCE LIBRARIES & DISTILLED MODELS
Production-grade open-source libraries, official framework contributions, and low-level neural network implementations.
Created a Keras 3-based tabular deep-learning library supporting JAX, PyTorch, and TensorFlow backends. Implemented state-of-the-art research architectures (SAINT, TabNet, TabTransformer, FT-Transformer, CTGAN, TVAE, GAIN, PCGAIN) for classification, regression, data imputation, and synthetic-data generation.
Contributed Beta and Binomial random-sampling functionality (keras.random.beta, keras.random.binomial) and multi-backend consistency fixes to the official Keras 3 API across JAX, PyTorch, and TensorFlow.
Inspired by GitHub Copilot, built a custom 260M-parameter contextual code-generation model based on Google DeepMind’s Gemma LLMs. Ported Gemma 2 from Flax Linen to the NNX API and applied staged knowledge distillation—first using Gemma 2 2B on text (FineWebEdu) and then CodeGemma 2B on fill-in-the-middle code (CodeContests) across 8× H100 GPUs.
Implemented Andrej Karpathy’s micrograd from scratch in pure C: a scalar reverse-mode autograd engine and neural-network library exposing a PyTorch-like API.
BUREAU DRAWER II // RESEARCH ENVIRONMENTS & HARNESSES
Open-source PyTorch evaluation harnesses and PettingZoo Parallel / Ray RLlib multi-agent simulation environments.
Open-source PyTorch implementation and reproducible evaluation harness for ADS-C, preserving every served top-1 prediction while reducing distilled-student accuracy by 29.6 pp on CIFAR-10, 17.4 pp on CIFAR-100, and 13.3 pp on Tiny-ImageNet across multiple student architectures.
Built a custom PettingZoo Parallel multi-agent environment with vectorized NumPy dynamics, energy-aware transitions, and proximity-triggered charging; trained parameter-shared IPPO/PPO policies with PyTorch and Ray RLlib. Established an analytical no-charge baseline and diagnosed a reward-hacking ‘lazy agent’ local optimum.
BUREAU DRAWER III // UNDERGRADUATE & EARLY ARCHIVE
Early work in computer vision, scikit-learn forecasting, Harvard CS50x, and NLP bots (2021–2023).