The problem
Most conversational interfaces are a thin layer over someone else’s model. That is fine in production and useless for understanding how the thing actually works.
What it does
Intent classification over a tokenised corpus, feeding a network trained in Keras, with the model serialised for reuse. Built to be inspected end to end.
The part that mattered
Owning the whole path from corpus to response. When it gets an intent wrong, the reason is visible in the training data rather than hidden behind an API.