What I'm building, breaking, and shipping. Consumer hardware, real experiments, real results. No corporate demos — just a person with a GPU and curiosity.
Agent proposes code changes to neural networks, trains on PubMed data, evaluates results, and iterates. Runs overnight on consumer GPUs with zero human intervention.
Result: 25% model improvement across 100 experiments. 93 failed. 7 stuck. The agent kept the 7 and built on them.
Hardware: RTX 3070, RTX 5070 Ti
Custom voice assistant built from scratch on Raspberry Pi 5 with USB mic and speaker. No Alexa, no cloud. Runs local models.
Result: Fully functional. Responds to voice commands, runs local inference, no cloud dependency.
Hardware: Raspberry Pi 5, USB peripherals
Multi-agent system where Claude agents coordinate to plan, research, implement, and review code changes. Powers this site's own development.
Result: Used to build and ship this entire site. Dogfooding the tools I sell.
Hardware: Cloud (Claude API, Supabase)
NVIDIA RTX 3070
ML training, overnight experiment runs
NVIDIA RTX 5070 Ti
Larger models, parallel experiments
Raspberry Pi 5
Edge AI, voice assistant, IoT
Cloud (Azure/Vercel/Supabase)
Production deployments, databases, APIs