What we're building, why it's shaped this way, and what we've learned along the way. Sovereign AI infrastructure from first principles.
We pointed 190 hostile prompts at our AI workspace and screenshotted every one. Every wrong action came from the deterministic regex router — at full confidence, in under three milliseconds — and none from a model. The bug that renames your document, why the precision of a good rule arms a bad one, why restraint is an anti-capability, and the three times we nearly fooled ourselves while measuring it.
A silent failure is the worst failure an agentic app can have: the user asks, the app accepts, and then nothing happens. How we hunted those ghosts across a 4,000-prompt corpus — the one missing word behind 96 routing leaks, a harness honest enough to tell a real ghost from a measurement artifact, and a hard rule that a good-looking number doesn't get to ship.
Two cheap wins that never touch the model: move fewer bytes, and don't ask the same question twice. Vera's transport compression (74% smaller responses) and per-tenant response cache (22× faster repeats) — the real numbers, the TTL model, and why tightening a PII rule invalidates the cache for free.
A user asked for a document and got a blank canvas; a 7am agent worked and told no one. The trace to the agent loop's last mile, the terminal contract we designed, the 100-prompt adversarial review that scored it at 31% — and the v2 that closed six structural clusters with fewer concepts: a manifest, a write-ahead journal, one reconciler, durable approvals, and a ledger for the irreversible.
A rigorous, honest investigation: thirteen models through a real replay harness, two embedders, four prompt levers, and a frontier adjudication panel. What actually moves a local-first router — and the four dead ends and one dangerous bug we found on the way.
A real 42k-photo library froze our photo grid. The investigation, the four fixes (thumbnails, a virtualizer, one SQL index, loading it all into memory), the self-inflicted 1,000-row bug — and what profiling the real library said.
We built a way to run other people's apps inside Veya, then wrote hostile ones and attacked the real build. Two holes, the fix — and the defense layer we were wrong about.
Veya the application fabric. Vera the intelligence hub. Open-weight models written for the work. Three layers, one stack, and the argument for why the public cloud can't ship any of them.
Why this blog exists, what to expect, and a note on shipping three platforms in a day.