I build assistants that work from a company's own logs, documents, tickets and code, and private model servers for data that can't leave the building. Each assistant gets read-only access, a budget and an audit trail before it gets its first user.
Before I set one up for you, I run one myself. This is the server behind my own assistants.
| Machine | Mac Studios with Apple silicon and large unified memory |
|---|---|
| Model | NVIDIA Nemotron 3 Super 120B-A12B (mixture of experts, 12B active) |
| Format | GGUF, UD-Q4_K_XL quantization, about 84 GB in memory |
| Server | llama.cpp llama-server, OpenAI-compatible API |
| Access | Cloud apps reach it through a locked-down reverse SSH tunnel. Nothing on the Mac is open to the internet. |
| Also on device | Apple MLX for task extraction in MyRunBook; Apple Foundation Models for the chat assistant in AstroStorm |
Answers product and support questions from app logs and a mirror of the source code. Read-only database role, one SELECT per query, no web access, per-user budgets and a full audit log.
Opened from inside an iPhone app, it lets a vendor ask why a badge or check-in failed. The company's server confirms who they are, and every tool is limited to that vendor's own records.
MyRunBook pulls action items out of meeting transcripts with a local MLX model, so transcripts never leave the Mac.
Claude Code, custom skills and MCP servers handle my App Store releases, nightly log digests, document filing and server upkeep. You get patterns already proven on real work.
I've taught technology courses since 2009. Today I run practical AI workshops for teams, built around their own tools and data.
Send me a short note on LinkedIn about what you're trying to do. We'll set up a free 30-minute call, and you'll come away knowing whether it's a fit, roughly what it takes, and what to do first.