How we build AI you own
We build AI systems our clients own outright — self-hosted, EU-compliant, and verifiable. This is not a product to buy; it is the method we prove on ourselves every day by running the whole company on it.
Six things we prove on ourselves
Every pillar is a real capability we run in production — not a slide. Where we have written it up, read the deep-dive.
We run open models (Gemma-class) on a consumer graphics card from 2018 — no cloud inference, no per-token bill, nothing leaving the machine.
Read the deep-diveWe fine-tuned a European-Portuguese guardrail that lifted PT-PT compliance from 87% on the base model to 100% — the kind of language quality generic cloud AI cannot guarantee.
Read the deep-diveA deterministic citation-guard fails any AI research note that cites a source the tool never actually opened. If it did not read it, it cannot cite it.
Read the deep-diveA repeatable 30-minute method to verify an AI-written plan before anyone acts on it — because unchecked AI output is a liability, not a shortcut.
Read the deep-diveThe whole company runs on a ~€7/mo VPS — Coolify, self-hosted Umami, Listmonk and SearXNG, a Tailscale mesh, encrypted R2 backups — €107/mo all in.
Read the deep-diveJUVENAL, our self-hosted agent harness, runs content, automation, infrastructure monitoring and client reporting — the method, working on ourselves every day.
Read the deep-diveLocal AI, proven on real work
The method is not theory. Here is our own local agent doing genuine, verifiable research on a real project — fully self-hosted, at zero cloud cost.
Bela Luz is our own permaculture homestead in Portugal — and the field-test where our local AI earned its keep. Running fully offline on one 2018 consumer GPU at zero cloud cost, pi built and maintains a source-verified knowledge base for the project: every fact checked against official Portuguese and EU sources — DGAV, EUR-Lex, the Diário da República — and when a claim could not be verified, the agent retracted it itself.
The agent is called pi. pi is our own local autonomous agent — the reference build behind the AI systems we deliver. It runs open models on llama.cpp, searches the web through our self-hosted SearXNG, and coordinates scout, researcher and verifier subagents — entirely on our own hardware, with no US cloud in the loop.
Read the deep-diveBela Luz is a personal project we use as a capability proof — not a client engagement, and the homestead itself is still pre-launch. What it demonstrates is the AI: a self-hosted agent doing real, source-verified research on infrastructure we own.
Why this matters for you
The same method, applied to your business, means every one of these is true of what we build for you.
You own the code and the infrastructure
Full handover of the system and the servers it runs on — inspect it, run it, keep it even if we ever part ways.
Your own operations dashboard
You watch the work on the same self-hosted KPI dashboard we run our own business on — no external SaaS, a weekly branded report, all on infrastructure you own.
No data leaves your servers
Self-hosted models and infrastructure mean client and business data never travels to a third-party cloud.
GDPR & EU AI Act by design
Compliance built in from day one on European infrastructure — not bolted on before an audit.
No vendor lock-in
Open-source foundations throughout — no proprietary control panel, no subscription you cannot leave.
No hallucinated outputs
The same verification and citation-guard discipline we run on ourselves protects the work we deliver to you.
Want AI you actually own?
Tell us the workflow you want to automate. We will show you exactly how we would build it — and how you keep it.