Crumb
Your store, now with a brain.
Created on 28th February 2026
•
Crumb
Your store, now with a brain.
The problem Crumb solves
The average e-commerce store loses over 70% of visitors before they ever add a single item to their cart. Not because the products are wrong — but because discovery is broken. Users who can't immediately find what they want don't dig deeper. They leave.
Crumb fixes this at the infrastructure level.
Our CLI plugin gives any online business — boutique retailers, large catalogues, food brands, specialty shops — a fully conversational AI agent that plugs directly into their existing stack. No storefront rebuild. No new backend. Just run the CLI, connect your catalogue, and your store now talks back. Customers can describe what they need in plain language, ask about ingredients, variants, or availability, and get answered instantly. The agent doesn't just respond — it guides, upsells, and closes.
We proved this end-to-end by deploying Crumb on a real bakery storefront. Customers could ask "do you have anything gluten-free under $10?" or "what's your most popular item this week?" and receive intelligent, inventory-aware answers in real time via voice or chat. The same agent that served customers also powered the admin layer — drafting new product listings, monitoring stock levels, and flagging high-intent conversations for operator review. One plugin. Both sides of the business, handled.
The discount negotiation flow alone demonstrates the commercial power of Crumb: a customer asks for a deal mid-conversation, the agent escalates to a live admin for approval, and resumes the conversation with the updated price — all without the customer ever feeling the seam. That kind of dynamic, human-in-the-loop commerce is simply not possible with static storefronts. With Crumb's CLI, it becomes a five-minute integration.
Challenges we ran into
The steepest challenge wasn't a single bug — it was the learning curve of orchestrating a real-time AI voice pipeline with stateful backend logic, something none of us had built before. VAPI's tool-call architecture operates statelessly: each function call from the voice agent is an isolated HTTP request with no built-in awareness of what happened before it. But our discount approval flow required state — the agent needed to pause mid-conversation, trigger an admin notification, poll for a human decision, and then resume with context fully intact.
Understanding how to architect that bridge — mapping ephemeral voice session context onto persistent database records, designing polling endpoints that wouldn't time out under VAPI's strict response windows, and handling edge cases where admins approved or rejected after the customer had already moved on — took significant iteration. We had to deeply internalize how VAPI's turn-based response model worked before we could reliably bend it to support a genuinely asynchronous human-in-the-loop moment. Once that architecture clicked, it unlocked the most compelling part of Crumb: an AI that doesn't just answer, but negotiates — in real time, with a real human backstop.
Tracks Applied (2)
Gemini API
Major League Hacking
Auth0
Major League Hacking