Each one is a small, public version of a problem companies actually have: answering questions over private documents, letting a model act on real systems safely, and not paying a frontier model for work a small one can do. The models are open and self-hosted through Ollama. Nothing you type is sent to a third-party model API unless the router decides it has to, and it tells you when it does.
Ask questions about a document and get answers with citations. Embedding, retrieval and generation all run on this server, so the document never leaves it. Use a sample handbook or, with a verified account, upload your own PDF. Uploads are deleted after an hour.
A Model Context Protocol agent wired to real tools, including a read-only PostgreSQL database, scoped per business scenario by an allowlist. Every tool call, its arguments, its result and any failure stream to the screen as they happen, so you can see exactly what the agent did.
Each prompt is classified for complexity, sensitivity and latency, then sent to the cheapest model that can handle it: local inference where possible, a cloud model only where needed. The decision and the money it saved are shown per request and on a running dashboard.
One account works across all three. Use the shared demo account to look around, or sign up with an email code for full access.
Open models served by Ollama on the same VPS, kept warm so the first answer is not a cold start. One request runs at a time; the rest queue and see their position.
NestJS APIs and Angular frontends in one pnpm workspace, sharing their sign-in, streaming and layout code. Qdrant for vectors, PostgreSQL for the agent's data, the router's history and accounts.
A small identity service issues signed tokens; each demo only verifies them against its published keys, so none of them can mint a session. Rate limits are per visitor.
Deepak Kumar Jha: Technical Lead · Solution Architect · AI Engineer, with 15 years of building and running production systems, based in Dubai.