Savia.

The problem it solves is old and expensive: in almost any company there are two or three people who know everything, and when they are away, everyone else waits. Savia takes what already exists — manuals, spreadsheets, meeting recordings, whatever gets answered over chat — and turns it into something you can ask in plain words. Every answer arrives with its source next to it, and when there is not enough material to answer, it says so instead of filling the gap with something that merely sounds right.

There are two pieces I am especially proud of. First: people data — salaries, ID numbers, sick leave — is not kept where the AI searches, so it cannot show up in an answer by accident; permission is checked on the server value by value, and something sensitive can be restricted to specific individuals. Second: when someone asks about something that is not covered, Savia notes it down, groups similar questions and, once several people have hit the same wall, sends a form to whoever does know. That answer gets reviewed, joins the knowledge base and notifies whoever asked. It is what makes the knowledge grow by itself instead of dying like a wiki.

Under the hood it is the most complete thing I have built, in case anyone cares about the detail. Every client company has its data separated at two independent levels, and inside each one knowledge is separated again by department, with automated tests verifying it on every change. Videos transcribe themselves and citations jump to the exact minute. It is built with FastAPI and Next.js on PostgreSQL and Qdrant; on my machine it runs on local models on the GPU and in production on the Claude API, recording what every query costs so I know the margin on each client.

Savia has been in production since July 2026 and keeps growing: it runs on my own server, with daily backups in a different region, alerts when something breaks, and complete verifiable deletion of a company's data the day they ask for it. The public site is at saviaia.es. It is sold by talking, not from a catalogue: first understand what hurts, then get it running in thirty days and stay for two more months until the team uses it without thinking. It is the project where I have learned the most about architecture, security and product.