← ALL WORK · CASE 11
WINSTON
A shopping assistant that understands what you're asking for — conversational product discovery with recommendations that learn from behaviour, across web and mobile.
DISCOVERY STATE MACHINE — EVERY QUERY WALKS THIS PATH
INTENT UNCLEAR → CLARIFYING QUESTION → REFINED MATCH — NO GUESSING AT CHECKOUT
01 — WARP
What had to be true first
Search boxes and filter trees make users do the retrieval work themselves. Winston's warp was an intent layer that takes it back: RASA for conversation management and intent recognition, TensorFlow for the recommendation models, and a Node.js backend with JWT authentication holding the whole thing together.
02 — WEFT
What we wove through it
On that base went the product: chat-based discovery where a shopper describes what they want instead of guessing the right filter, recommendations built from actual behaviour rather than category averages, and Phoenix webhooks carrying real-time events so the assistant reacts while the user is still looking.
03 — BEAT
How it was pressed tight
Conversational systems fail politely, which makes their failures easy to ship. The beat concentrated there: ambiguous intents get a clarifying question instead of a confident wrong answer, recommendation quality is measured against real sessions, and the event pipeline is idempotent so a replayed webhook can't double an action.
04 — WEAR
Life in production
Winston runs across web and mobile with product discovery that feels like asking a knowledgeable shop assistant rather than operating a database. Engagement went up because the effort of finding things went down — which was the entire thesis.
NEXT CASE — 12
MANIYARI →