I build my own AI and hardware systems to pressure-test strategy before scaling it to teams.
Shaw AI
Self-hosted multi-agent platform, 2026 to present
- A local-first agentic AI platform across 6+ nodes (Apple Silicon inference, Linux canary, Raspberry Pi edge) on a Tailscale mesh, running two isolated agents with least-privilege access to a shared Obsidian knowledge vault.
- Moved the primary agent from a cloud LLM API to on-prem open-weight inference: idle API spend from about $5 a day to $0, 10+ models evaluated on tool-call reliability and latency, Ollama tuned to about 52 tokens per second, and per-turn context cut about 80% (67K to 14K tokens).
- Run like production: canary-before-promote releases, version pinning, incident RCAs and runbooks, and a verbatim-output rule for agent reports after catching an agent fabricating a completed check.
GregOS
Python, YAML, local vision models
An ingestion pipeline with inbox, processing, completed, and failed states that turns PDFs, images, handwritten notes, and phone uploads into structured Markdown and auto-generated slide decks. Covered by unit, integration, and smoke tests.
Meeting intelligence and morning brief
iOS Shortcuts, Python, TypeScript, local LLM
A privacy-first service that summarizes long transcripts with map-reduce on a local model, and an agent that compiles email, calendars, and notes into a daily executive brief. Both run unattended.
StanleyFace
ESP32-S3, C++, Flask, Whisper, streaming TTS
A multi-persona voice AI robot where adding a character is a config change. Streaming speech reaches the first spoken word in about one second.
40 Hz sync device
ATmega328P firmware, custom 2-layer PCB
Designed, fabricated, and validated a device that optically locks to a 40 Hz light panel and plays a phase-synced tone: about 41.6 Hz lock, 1 ms jitter, zero dropouts over 15 minutes.