AI-native Products
Two builds where AI runs the whole experience instead of sitting off to the side as a feature. One is a Shopify operations agent. The other reimagines Airbnb as something that plans your trip for you.
Kandwii
github.com/adam7rans/shopify-agentAn AI operations agent for a Shopify store. I built it as my AI engineering test for Growth Capital.
Kandwii is an AI agent that helps Shopify store owners run their store by talking to it directly, no dashboards or reports required. Ask it a question and it figures out which tools to call, pulls the real numbers from the Shopify Admin GraphQL API, and answers with whatever fits best: a card, a table, a chart, an invoice breakdown, or a block of Shopify Liquid. There is no hardcoded list of question types behind it. In this build, the store is an Asian candy shop called Kandwii, used here as the demo storefront.
What it does
- Sales analytics. Best sellers, revenue, margin, and category breakdowns, shown as charts.
- Inventory and reorder. Stock levels, plus stockout risk and reorder suggestions based on how fast things sell.
- Warehouse health. 3PL snapshots that flag delays, stuck fulfillments, and damaged stock.
- Document handling. Reads invoices with OpenAI vision, works out the inventory impact, and drafts the reply to the supplier.
- Liquid generation. Writes Shopify collection and page templates from your actual products.
- Proof mode. A live activity log so a reviewer can watch the real tool calls and API requests happen.
How it works
A prompt goes through caching, an OpenAI tool-calling loop, and a response validator before it streams back as UI. Click through the steps below to see a request move through the system.
Click a step to walk through the flow.
One agent loop, not a switchboard
Everything runs through a single OpenAI agent loop. There are no fixed routes for each kind of question. It always reads the real store data before it answers and it never makes numbers up. A mock and live Shopify Admin GraphQL setup lets a reviewer run it against real data or seeded data.
The stack is Next.js, React, and Tailwind on the front end, OpenAI tool calling for the agent, the Shopify Admin GraphQL API for commerce, Convex for saving conversations, and Recharts for the graphs.
AI-Airbnb
github.com/adam7rans/AI-native-airbnbA detailed prototype of what Airbnb could be if it planned the trip with you. I call it the Adaptive Trip Canvas.
View it live → ai-native-airbnb.vercel.appRight now Airbnb opens with an empty box that asks “Where are you going?” This version opens already knowing. It looks at how you travel over the years, the bookings you made, the homes you saved, the photos you lingered on, the time of year you usually go, and it hands you a trip that is already half planned. The tone is closer to “I think I know the trip you’re after. Here’s a start. Tell me what to change.”
Key flows
- Adaptive Trip Canvas. A home screen it puts together for you instead of a search form.
- Live Trip Brief. A running summary of what it thinks you want, which you can edit by voice, typing, taps, or gestures.
- Repeat last year. Rebuilds last year’s home, experience, and services and checks if they are free.
- Same trip, but better. Say “quieter this year” and it adjusts, then tells you why it changed things.
- Draw to search. Circle part of a photo and talk, and it turns what you drew and said into a real search.
- Itinerary remix. Reshuffle a multi-day plan just by talking. Move things, drop things, add things.
How far it goes
This is a clickable storyboard, not a shipping app. It is a set of phone-sized frames that walk through each flow step by step, from opening the app to tapping, scrolling, talking, drawing, and seeing the result, built with React, Vite, and Tailwind. It also comes with a PRD and a backend write-up covering travel memory, vector search, a trip knowledge graph, and GraphRAG, so the path to a real version is on paper too.