Inspiration
Second-hand stores like our beloved Ithaca ReUse face a unique problem: every donated item is different. Traditional inventory systems expect standardized products with UPCs, SKUs, and consistent names, but ReUse gets everything from curtain rod brackets to oddball hardware pieces to unlabeled electronics.
Volunteers told us they can only categorize items into broad buckets like Hardware, Clothing, or E-waste, which makes the inventory nearly impossible for customers to search meaningfully. We wanted to build a tool that actually understands second-hand items and makes them findable.
What We Built
We created ReFind AI, an AI-powered intake and search system designed specifically for second-hand environments. It has two main components:
- AI-Powered Donation Intake
Volunteers upload a photo → the system automatically generates: • a standardized name • a fine-grained category • tags • an item description
This transforms the chaotic diversity of second-hand items into structured, searchable data.
- Semantic Customer Search
Customers can search using natural language like:
“Do you have the little thing that holds up a curtain rod?”
Even vague descriptions return the correct items.
How We Built It
• Frontend: React Native
• Database: Supabase / PostgreSQL with item metadata + vector embeddings
• AI Models:
• Vision model for object identification
• LLM for naming, categorization, and descriptions
• Embedding model for semantic search
Challenges faced
- Categorizing Non-standard Items
Unlike retail items, second-hand goods have no consistent taxonomy. We had to build prompts and logic that create hierarchical, consistent, and meaningful categories.
- Avoiding Over- or Under-Categorization
Early versions produced categories that were too generic (“Hardware”) or too specific (“Vintage 1997 Heavy-Duty Steel Wall-Mount Bracket Model X”). We iterated heavily to strike a balance.
What We Learned
• How different the data model of a second-hand store is compared to retail
• The importance of consistent taxonomy for inventory health
• How to combine vision models + LLMs + database
• How challenging it is to design usable AI tools for real volunteers and shoppers
• That even “simple” items like brackets or adapters require careful metadata design to be discoverable
Built With
- javascript
- reactnative
Log in or sign up for Devpost to join the conversation.