English · Deutsch
A German craftsman running his solar business on local AI. Two NVIDIA DGX Sparks, one Mac Studio, real customers. Everything reproducible, every source credited.
I'm Julian, I run a solar installation company in Freiburg, Germany (8 people, 1,000+ installs). I'm not an ML engineer. This repo is my lab notebook: what I run, what I measure, what breaks.
→ Follow along on X: @HandwerkAiLab
| Component | Role |
|---|---|
| 2× NVIDIA DGX Spark (128 GB unified memory each) | Inference servers (vLLM / SGLang) |
| Mac Studio | Local agent host (HERMES), MLX experiments |
Mellanox ConnectX NIC in OWC Mercury Helios 5S (Thunderbolt) – model being verified, see hardware/ |
RDMA link Mac ↔ Spark via MCDMA |
| DAC cable (type depends on NIC ports) | Physical link Mac ↔ Spark |
Details: hardware/
| Folder | Content |
|---|---|
logs/ |
Lab log, one entry per day/week, each linked to its X post |
recipes/ |
Model setups per Spark configuration |
benchmarks/ |
Measurements with date, method and raw data |
agent/ |
Agent config templates (no keys, no customer data) |
mcdma/ |
Notes on the Mac ↔ Spark RDMA setup |
notes/ |
Field notes: DGX Spark networking, known issues, what makes it fast |
SOURCES.md |
Every repo, post and paper I built on |
- Local first. Customer data never leaves the building.
- Human in the loop. The AI drafts, a human sends. Nothing is published or sent automatically.
- Measured vs. estimated. Every number states hardware, model, quantization and date.
- Credit. If I used your work, you're in
SOURCES.mdand tagged in the post.
Scripts: MIT · Texts: CC BY 4.0 · Third-party recipes keep their original license.