mySommelier. A wine recommender that actually learns.
Five LCBO Vintages bottles you'd love this Saturday, picked by an agent that updates its taste model every time you rate a wine.
The bet.Wine recommenders fail because they don't persist taste memory. A profile that recomputes from every loved and disliked wine, plus a fresh local catalogue, beats both generic LLM chat and static review apps.
The product call. Picks first, chat as refinement layer. Conversational onboarding cut for a 5-question form plus 2 to 3 reference wines. Public profile pages with no login wall on the demo, because the product is the recommendation.
The eval work.I caught the model being confidently wrong four ways during build. Suggesting a white when I'd only logged reds. Picking a 14.5% bold when I'd said medium. Drifting above my sugar ceiling. So the system ships with three eval layers: a data-quality eval upstream of everything, a retrieval eval with four labeled personas and a hand-coded ground-truth function (NDCG@5, recall, diversity), and append-only trace logging for L2 review. Provider-agnostic, so swapping Gemini for OpenAI is a config change and the harness re-runs against the new embedding space.
What I cut. Multi-store inventory (the job is “should I buy this,” not “where is it”). Free-text NLP onboarding (kept as a bet, not an assumption). Login walls on the demo.