Inspiration
Software track
People want shopping help that feels human: quick product discovery, follow-up questions, and recommendations that understand context and taste. Silktree AI was inspired by the gap between static e-commerce filters and a friendly, guided shopping conversation — an assistant that can browse, clarify, and recommend across multiple turns. And it can be integrated with any website UI so navigation is same whether you are on Amazon, ebay or Macys.
What it does
Silktree AI is a conversational shopping agent that helps users find products through natural language. It understands intent, asks clarifying questions when needed, filters and ranks relevant items, remembers session context across turns, and surfaces concise product suggestions with reasons and links. The frontend demo lets users chat and receive product cards and suggestions in real time and make payment using Agent wallet.
How we built it
We used a modular agent architecture: a lightweight Python backend that coordinates specialized agents (search/scout, product checkout, eye/vision or enrichment, and an orchestrator) and a session manager to keep multi-turn state. The frontend is a minimal chat UI that sends utterances and renders product suggestions. Core pieces include intent parsing, retrieval from a product index, simple ranking heuristics, and explicit context passing between agents for safe, explainable responses.
Challenges we ran into
Maintaining coherent multi-turn context and deciding when to ask a question versus when to recommend was tricky. We tried to do checkout using chat, and it thought that we want to buy some checkout ring, and we were having limited token per day so we couldn't test it for all the edge cases.
Accomplishments that we're proud of
We built a working end-to-end prototype that can hold multi-turn shopping conversations and produce useful, explainable suggestions. The modular agent design made it easy to swap or extend capabilities We also shipped a lightweight frontend that demonstrates the user flow clearly and is ready for user testing.
What we learned
Grounding recommendations in verifiable product attributes improves trust more than opaque high-scoring suggestions. Early, simple clarifying questions dramatically improve match quality. Architecting for small, composable agents makes iteration faster and testing easier than a single monolithic model. .
What's next for Silktree AI — Conversational Shopping Agent
Short-term: integrate richer product metadata and an embeddable product-index for semantic search, add metrics (conversion/engagement) and an A/B test harness, and tighten session continuity. Medium-term: add personalization, a small recommender model, and a controlled rollout to real users for qualitative feedback. Long-term: support multimodal queries (images + text), deeper cart/checkout flows, and store integrations for live inventory and pricing.
Built With
- gemini
- javascript
- langchain
- mcpui
- python
Log in or sign up for Devpost to join the conversation.