Project : mentat

Ask a general model a product-strategy question and you get confident prose with no ground under it. mentat answers the same question and shows you where each part came from.

It is a two-service application for European HR-tech questions. Retrieval runs keyword and meaning search together and fuses the two rankings, because short queries break one and paraphrase breaks the other. Four tools sit on top: SWOT, RICE scoring, competitor comparison, and live web search. A safety gate sees every input first.

The instance it ships with advises a fictional Berlin HR company, so the work can be judged without a real client’s data in it.

Citing is not decoration. It is the only thing that makes an answer checkable.

What it is made of

An HR-tech strategy advisor from my 2026 Turing College AI Engineering programme, on hybrid retrieval and an evaluated answer set. One of the three products I shipped in that programme, each with its own evaluation harness.