What we were handed.
Building AI products on rented APIs means per-token bills, rate limits, and data leaving the building on every request. That's fine for a prototype; it's a poor foundation for products that run around the clock.
We wanted to build AI features where the economics and the privacy are ours — and where an experiment that runs all night doesn't show up as an invoice.
The system.
A dedicated AI compute node serving open-weight language models behind one API, plus an image-generation studio, speech-to-text and text-to-speech, and a private meta-search engine.
On top of it, a hybrid routing layer: local models handle the volume, and frontier cloud models are pulled in only for the tasks that genuinely need them. Running it for real also taught us the failure modes — the stack now has memory guardrails so one heavy job can't take the box down.
Open-weight models behind one consistent API for every product.
Generation, transcription, and voice — all on-prem.
A router that picks local vs cloud per task, so cost follows value.
Memory and load protection earned the hard way, in production.
What changed.
Every AI feature we ship — Voxelburg's residents, the ops fleet, the assistants — runs on this stack. Fixed costs, no rate-limit surprises, and data that stays home by default.
The next study could be yours.
Bring us the problem — the messy, manual, “there has to be a better way” kind. We'll design the system, build it, and run it.