Conversation as the interface for an inventory where nothing repeats

We positioned conversation as the primary interface rather than an assistant added alongside a conventional storefront. The distinction matters commercially: a chat widget bolted to a filter-and-grid site is a support channel, while a conversational agent that holds intent is the mechanism by which a one-of-one inventory becomes searchable at all.

Describing what you want in your own words — a brand, an occasion, a silhouette, a budget — suits resale better than a filter set, because a shopper's intent is usually broader than any single listing and narrower than any category. The same reasoning applies to supply: if a seller can describe an item rather than complete a form, the cost of listing drops to something a casual seller will actually pay, and the catalogue grows.

A shopper describes what they are looking for, the agent interprets it against live inventory, and results are returned as a considered set with the reasoning visible — why these pieces, and where a compromise has been made on size, colour or price.

Where nothing matches, the interaction does not end. Intent persists, so an item listed later can be matched back to a shopper who asked for it days earlier. Conversation is not the only way through: results resolve into familiar product and checkout surfaces, and a shopper who prefers to browse can do so — the agent is the fastest route to a match, not a toll gate in front of the catalogue.

Listing was designed as the shortest possible path from a garment in someone's hands to a live, searchable item. A seller photographs and describes; the structure a marketplace needs — category, brand, size, condition, measurements, price guidance — is assembled around that rather than demanded upfront.

This is as much a data problem as an experience one. Search quality depends entirely on listing quality, and casual sellers cannot be relied on to categorise consistently. Generating structure from natural description keeps the catalogue coherent while keeping the effort low enough that people finish. Seller-facing surfaces were kept deliberately plain — someone selling three items from a wardrobe should not be given a merchant dashboard.

Second-hand buying asks for confidence a new-goods marketplace can assume. The buyer is judging condition from photographs, sizing from a brand whose fit is inconsistent, and a seller they have never dealt with — on an item that cannot be replaced if it disappoints.

Condition, measurements and seller standing are presented consistently on every listing so they can be compared across items, and payment and delivery are handled by the platform so neither party is asked to trust the other with money. The agent contributes too: a shopper can ask about fit or condition in the same conversation in which they are choosing — the point at which the question actually arises.

The platform was built to launch as a working two-sided marketplace rather than a prototype — listing, discovery, checkout, payments and delivery in place from day one, since a resale marketplace cannot be tested honestly without both sides running.

Interface components were kept restrained on purpose. In a catalogue where every item looks different, the surrounding system has to stay quiet or the page becomes noise. The agent's behaviour is treated as a design surface that will keep changing — how it asks, how much it assumes, when it hands back to browsing — tuned against real conversations without rebuilding the experience around them.

Thrifti launched in 2026 as a functioning circular fashion marketplace for India, with conversation as its primary route to a match rather than an addition to a conventional storefront.

The proposition it launched on — keeping clothing in rotation, and making a one-of-one inventory genuinely searchable — is now live and being tested against real supply and real demand. Performance data will follow the launch period.