The problem
Category trees and keyword search were a 1995 idea. People don't shop by SKU; they shop by outcome — "a winter jacket that doesn't look like every other winter jacket," "a gift for my dad who likes fishing," "headphones for my commute under 1500 TL." The current commerce stack forces them to translate that intent into a search query, then into a filter, then into a comparison spreadsheet of their own making.
Most stores respond by adding a chatbot to a category page. That's a worse search bar with a personality. The shape is wrong.
Our angle
Conversational discovery as a first-class shopping primitive — not a chatbot bolted onto a category page. The buyer describes the outcome. The platform interprets, retrieves, and refines. The output is not a list of links; it's a conversation that ends with the buyer pointed at the right product across the right store.
The substrate is the same one Atlas and Apollo run on. Catalog facts are normalized. Pricing is current. Stock is real. Recommendation isn't a black-box model; it's a structured query the platform composes against a federated index — and the buyer can see the reasoning at any depth they care about.
What we're exploring
Embedding-based product search at speed and across stores. Conversational refinement loops that don't lose the buyer's intent across turns. The trust signal when the platform is uncertain — "I have three good answers; here's the one I'd choose, and here's why." Cross-store recommendation when no single store has what the buyer wants. The right ratio of generated text to product imagery in the response.
Latency is the brutal constraint. A buyer who waits two seconds is no longer in the conversation.
Status
Early research. Internal prototypes against a small catalog. Not customer-facing. No timeline; we will say when, plainly.
An invitation
If you've worked on retrieval-augmented generation, embedding indexes at scale, or the shopping-UX layer of recommendation systems, we want to hear from you. research@fastart.tech.