AI to empower people in need
and those who help them

The agent answers from your handbooks, protocols and official references: resource libraries, tagged documents, PDF import, a site indexed by crawl. Never from the open web.

The sources panel shows the document title and the extract used, so whoever reads the answer can check it. An answer nobody can verify has no place in a public service.

The agent's settings are your organisation's. Every reply has to report its own bookkeeping: session title, categories drawn from your list, sources cited. The platform enforces that when the model skips it.

Every change writes a draft. Publishing makes it the version that runs, named and dated in the history, and every reply carries its version badge. Nothing reaches the public that a human has not published.

Evaluation campaigns replay conversation and extraction datasets against a chosen version. Reviewers rate the answers blind, an LLM judge gives the scale, and two runs compare side by side.

Every exchange is filed under the life situations you defined. The topics your answers cover badly surface on their own, instead of waiting for a complaint.

Each agent exposes itself as an MCP server, so the assistant or portal your institution already runs calls it as a tool. A widget on your site or a public page when that is simpler. Nobody changes screen.

Managed cloud, or your own infrastructure from a single Docker image. Document isolation is written into the retrieval query itself rather than a checkbox, the retention period is yours, and the code is MIT-licensed.