What we build
Sovereign AI systems: they run on hardware you own, and the data never leaves it. What that buys, what it costs, and the numbers.
1 Nov 2026 · 1 min readSovereign AI systems — they run on hardware you own, and the data never leaves it.
What sovereign means here
Not: a cloud API with a privacy policy. Is: a 12-billion-parameter model doing real agentic work on one machine, with no inference call leaving the building.
What that buys:
- No per-token bill that scales with success
- No vendor deprecating your system
- No compliance question about where the data went
What it costs: engineering. A cloud call is one line. This is not.
Measured, from the system we run
| native tools | 69 |
| TypeScript | 64,644 lines on 29 runtime dependencies |
| written by hand instead of imported | 2,005 lines, replacing 12 dependencies |
| retrieval over 6,346 facts | ~150 ms |
| custom inference runtime | 2,841 lines of Rust |
Twenty-nine dependencies is the number worth looking at twice. Each one removed is a supply chain you no longer inherit.
One result that changed the architecture
We asked the system to retrieve a behavioural rule.
- Stored as the rule’s own words: found 1 time in 5.
- Stored as the situations that trigger it: found 10 times out of 10.
The technique is weaker on tool selection — 7 out of 10. We report that too.
Knowing where your own method stops is part of the method.
The offer
Custom AI systems: sovereign, cloud, or both where it makes sense. Consultations and workshops.
Base rate 100 €/h. All three need preparation and research, reported with the artifacts produced and charged at 50% of base rate.
Non-commercial project. The offer exists because the work has to be paid for, not because this is a funnel.
Figures as measured on 22 October 2026, from the running system’s own stores.
Gemma-4 26B swapped 27,848 times in a real agentic run. One flag reported zero, on an easy test. Under the load we actually run, it took two settings together: experts on the CPU and an 8-bit KV cache. No new runtime needed.
Our local agent took 24 seconds to say 'OK'. The loop was fine. A cache flag was set to zero, in three places, and the one we edited first was not the live one.
On a paid API, prompt size is the enemy. On a local model with a prompt cache, volatility is. A 21,900-token per-turn bill, where the tool schemas were the bigger half, and the one ordering rule that serves both.