Welcome to my little corner of the internet! I'm your local tech bro — fifteen years of shipping, a founder who builds and still runs his own profitable products, and the person people have always brought their hardest systems problems to. Uncle Sam included.
The newest chapter is ML infrastructure. I'm putting five years of locked-down AWS work toward making machine-learning systems cheap and reliable to run: benchmarking LLM inference on Apple Silicon, building zero-touch MLOps pipelines, and fine-tuning domain models locally with MLX. Same systems brain, pointed at inference cost. If your inference bill is outgrowing your revenue, that's the person right in front of me I like to solve for. Otherwise, I'm probably deep in a Sudoku.
Building and running my own products (VisorPlate, Widda) while adding an ML-infrastructure edge: a public inference benchmark suite, an automated MLOps pipeline, and a locally fine-tuned domain model — each with real, reproducible numbers. Same systems judgment, now pointed at inference cost.
Five years building secure, AWS-hosted systems for government and enterprise. Owned CI/CD, infra hardening, and security review across parallel production projects — the systems foundation the ML work is built on.
Full-stack engineer who drove a stalled iOS product to launch. The company was acquired by DraftKings in 2019.
Rigorous performance analysis of open-source LLMs on Apple Silicon — comparing quantization strategies, memory throughput, and tokens/sec across Llama 3.2, Mistral 7B, Phi-3, and Qwen2.5. A clean CLI anyone can run, built on MLX and llama.cpp.
↳ fp16 · q8 · q4 · q2 — reproducible latency, memory & quality deltas
A fully automated ML pipeline — ingestion, training, evaluation, versioning, and deployment — built with MLflow, Prefect, and AWS. An evaluation gate blocks any model that regresses. Zero manual steps from raw data to a live FastAPI endpoint.
↳ data → train → eval-gate → versioned artifact → live endpoint
Fine-tuned Mistral 7B on a niche, publicly publishable dataset using MLX-LM, natively on Apple Silicon. The differentiator is the custom evaluation harness: task-specific test cases measuring the delta against the base model — published, not vibes.
↳ published to Hugging Face Hub · measured before/after quality delta
AWS Certified Machine Learning Engineer — Associate
Amazon Web Services· role-based MLOps cert
Practical Deep Learning for Coders (Parts 1 & 2)
fast.ai · Jeremy Howard· 2026
Hugging Face NLP / LLM Course — + published model on HF Hub
Hugging Face· 2026 · fine-tuned model live with downloads