Data scientist and ML engineer focused on practical AI systems.

I focus on small-scale LLM training, evaluation, and shipping the full app around the model.

A decoder-only GPT-style transformer designed and pretrained from scratch in PyTorch on a single RTX 3090.

• Final 151.9M v3 model trained on a self-curated 10B-token dataset
• Beats GPT-2 small overall on a fixed eval suite: held-out perplexity, WikiText-2, LAMBADA, and multiple-choice continuation scoring
• Open source code, published Hugging Face model cards, and live completion demo

Code · Model collection · Final v3 model · Live demo

• Small / efficient models that punch above their weight
• Training tricks that matter on real hardware
• Honest model evaluation and benchmark comparisons
• End-to-end AI products — model in the middle, real product around it
• Local-first AI tooling on consumer GPUs

• Primary — Python, PyTorch, FastAPI, SQL
• ML / Data — Hugging Face (Transformers, Datasets), NumPy, pandas, Polars, scikit-learn
• Also — Node.js, R, Java, JavaScript, Streamlit
• Environments — Linux, Docker, Git

• UT Austin — Post Graduate Program in Generative AI for Business Applications (Jan 2026)
• Arizona State University — B.S. Data Science, summa cum laude (May 2024)
• SmartRent — Workforce Management Analyst (forecasting, scheduling optimization, KPI pipelines)

• Portfolio — bmax16634.github.io
• Hugging Face — @bmax16634