VP at IT Outsource · Marshall, TX · 🦞
Self-taught. Hands-on. Eight years of keeping 100+ businesses running on iron in a rack, not magic in the cloud.
Per-layer Residual VQ codebook over StyleGAN2 W+ space. 820× compression, 64-second training on an A4000. Compresses face representations to 180 bytes with minimal quality loss.
Self-hosted remote access dashboard wrapping MeshCentral. Built to replace ScreenConnect across 100+ client environments. Dogfooding on the homelab first.
Small-scale LLM training with GRPO reinforcement learning. Demonstrated that RL creates genuine capability at 10.7M parameters. Open-source experiment pipeline.
Building the Sovereign Stack — locally-hosted alternatives to the everything-as-a-subscription model. Email, VoIP, remote access, file services, all running on hardware I own.
Migrating clients off ScreenConnect onto MeshCentral. First one done, many to go.
Training small language models with RL. Turns out 10M parameters is enough to see real reasoning emerge.
Playing the tenor sax again. Might write a song. Might not. The important part is picking it up.
Trying to finish projects instead of starting them. So far so good.