Savia.

Technically it is the most complete thing I have built: a multi-tenant platform with FastAPI and Next.js where data isolation is twofold, RLS policies in PostgreSQL plus one Qdrant vector collection per company, with a second permission level per department. The RAG pipeline ingests documents and videos (transcribed automatically; every citation jumps to the exact second), chunks them, embeds them and retrieves with a confidence threshold: when the knowledge is not enough to answer, Savia says so instead of making things up. In development everything runs on local models on my own GPU, and in production on the Claude API, recording the cost of every call per company.

There are two pieces I am especially proud of. First, employee records: people data (ID numbers, salaries, sick leave) lives outside the AI retrieval so it can never leak through a lucky vector match; the chat queries it through tools whose permission is checked on the server on every call, down to the level where a sensitive field is visible only to specific authorised people. Second, the self-improvement loop: when someone asks something the knowledge base cannot answer, Savia records the gap, clusters similar questions and generates a form for the right expert; after review, the answer is indexed and the loop closes by notifying whoever asked. On top of that, whatever the team solves in the chat is captured automatically as reusable knowledge, and an internal panel runs the business side: approving companies, audited support and metering usage, costs and per-seat billing with Stripe.

Savia is still in active development and it is the project where I have learned the most about architecture, security and product. The public landing is already live at saviaia.es, with a consultative sales model aimed at SMBs: first understand the client's problem, then roll out the tool.