Emma AI

A grounded support assistant that answers from the client's own material - and nothing else.

$18,000 / 1,5 months

Python + FlaskAPI

Turso / libSQLnative vector search

OpenAI embeddings, Claudeanswers, reranking and grounding

React + Vitechat client

About this project

Emma AI is the built-in assistant of an e-learning company's coaching product. Participants used to wait on human support for questions the course material had already answered; Emma answers them in seconds, in German, in the tone of the brand. What makes it work is the control surface around the model: the client curates exactly what Emma is allowed to know - FAQs, course modules, PDFs, images - and every reply is retrieved, reranked and grounded against a source before it is sent.

The challenge

Hallucination turned out to be a product problem rather than a model problem, and the fix was structural: retrieve, rerank, ground against a cited source, and refuse when confidence is low - a refusal is a better answer than a plausible lie. Scope is enforced at retrieval time, so material a member has not paid for never enters the context window at all.