AI-native marketplace

Shopping by intent, not filters: an AI-native marketplace that answers a shopper's plain-language request with reasoned picks, and shows sellers what people are looking for.

All projects

Whose need

People think about what they need in sentences: "a rain-proof tent for two, under $300." Marketplaces still answer with category trees and filter checkboxes. The right product gets buried under thousands of results, finding a gift eats an evening, and a return turns into a form marathon. Sellers have it no better: listing a product takes time, and demand stays invisible.

What I built

I designed Askwell as a marketplace where AI isn't a side feature but the interface itself. The search box gives way to an intent box. Shoppers say what they want; Askwell searches, narrows down, compares and explains every pick. Gift finding, photo search, review summaries, virtual try-on and conversational returns are all part of the same experience. Sellers get one-step listings from a photo and an assistant that surfaces demand before it peaks. And Askwell is open to AI agents, not just people: through MCP and UCP, other assistants can shop here too.

The flow

The shopper writes what they need; Askwell breaks the sentence into meaningful tags.
Hybrid search and LLM reasoning scan thousands of products and return results with a "why this" for each.
Review summaries and a personal comparison settle the choice; virtual try-on removes the last doubt.
After purchase, tracking, returns and exchanges happen in conversation, and every action waits for the shopper's approval.
Sellers list a product from a photo; the assistant shows what people are searching for and what's missing from stock.

Key decisions

Intent instead of a search box
Generative discovery doesn't return a ranked list. It plans several searches, reasons over the results and recommends with reasons, so shoppers always know why they're seeing a product.
Conversation and storefront on one surface
Instead of a chat bubble squeezed into a corner, generative UI: the AI answers with product cards, comparison tables and return cards, not paragraphs. While the AI works, the interface dims; results flow back onto the light canvas. Shoppers can always tell when the AI is at work.
Every action that touches money or orders is approved
Returns, exchanges and checkout steps go through human-in-the-loop approval. Layered guardrails, prompt-injection defenses and an eval suite running in CI (golden datasets, LLM-as-judge) make every answer measurable.
A marketplace open to agents
Built for agentic commerce: with an MCP server and UCP/ACP support, a shopping agent in ChatGPT or Gemini can query the Askwell catalog in real time.