The brief
Make music discovery conversational. Instead of a user guessing the right keywords, let them describe what they want and have the platform understand it.
Why most AI features never ship
A demo works once, on a laptop, for a friendly input. Production is a different problem, and it has three parts.
Latency: an LLM call sitting in a user-facing search path has to return fast enough to still feel like search. If it takes four seconds, users stop using it — the feature is technically live and practically dead.
Cost: a call that costs a fraction of a cent is irrelevant in a demo and existential at a million queries. The cost model has to be worked out before you build, not after the first invoice.
Failure: models are slow sometimes, and wrong sometimes. A production feature needs a defined answer to "what does the user see when the model returns nothing useful" — and that answer cannot be a spinner.
Designing around those three, rather than around the happy path, is the whole difference between a demo and a feature.
The result
The platform shipped with LLM-driven conversational discovery — architecture designed end to end, and the model integration built into a real product rather than a proof of concept.