Spark Landing Kit
Two DGX Sparks to private AI in a day.
A deployment kit that turns two NVIDIA DGX Spark computers into a private, OpenAI-compatible AI service in one day. Every step ends with a pass/fail check, so nothing touches sensitive financial work until it's proven.
- Promotion gates
- 8
→One OpenAI-compatible gateway, three serving lanes
→Eight promotion gates from license check to long-context needle tests
→Model-agnostic lanes: any checkpoint meeting the contract drops in
Two ways to read it
In plain English
A deployment kit that turns two NVIDIA DGX Spark computers into a private, OpenAI-compatible AI service in one day. Every step ends with a pass/fail check, so nothing touches sensitive financial work until it's proven.
For engineers
Bash + stdlib Python, LiteLLM gateway routing to three vLLM serving lanes over Tailscale, with eight promotion gates (license, boot evidence, warm decode, behavior, long-context needle tests).
What it does.
- One OpenAI-compatible gateway, three serving lanes
- Eight promotion gates from license check to long-context needle tests
- Model-agnostic lanes: any checkpoint meeting the contract drops in
Built with
- Bash
- Python
- LiteLLM
- vLLM
- Tailscale
- DGX Spark
Status: In progress · started July 2026 · open on GitHub