Stockpulse

Inspiration

Small shop owners run their inventory the same way they did decades ago, checking shelves by eye and finding out they're out of stock only when a customer asks. That means lost sales every single week, and no affordable tool built for their actual scale. I wanted to give small business owners the same kind of inventory intelligence large retailers have, without needing enterprise software or a team to run it.

What it does

StockPulse is an AI powered inventory reorder assistant built for small shop owners. A user takes a photo of their shelves, and the Gemini API analyzes the image, identifies each item, and estimates quantities, with no manual data entry and no barcode scanning required. StockPulse then tracks how inventory depletes over time, and when something is about to run low, it automatically sends a reorder alert, along with supplier suggestions, so the owner has a clear next action instead of just a warning.

How we built it

The Gemini API powers vision based extraction, reading a shelf photo and converting it into structured item and quantity data. Supabase, running on Postgres, stores inventory, depletion history, and reorder data. Firebase Authentication handles login through Google Sign In and email and password. The app is deployed on Vercel. Almost all of the code was written through vibe coding using Google Antigravity, an AI coding agent, since I can't code much myself.

Challenges we ran into

Getting Gemini's vision extraction to reliably identify individual items and reasonable quantities from a single photo, rather than a rough guess. Structuring depletion tracking so reorder alerts actually reflect real consumption patterns instead of a flat threshold. Learning to work almost entirely through an AI coding agent rather than writing code directly, which meant getting good at describing exactly what needed to change and verifying it actually worked once deployed, rather than assuming a reported fix was correct.

Accomplishments that we're proud of

Going from an idea to a real, working, deployed product, built almost entirely through AI assisted development, with real users already using it to track their own inventory.

What we learned

Building StockPulse taught us that AI vision extraction needs real world testing, not just a working demo case, since actual shelves are messier and more varied than a single test photo. We also learned that relying on an AI coding agent for nearly all development changes how you work. You have to verify results yourself rather than trust that a reported fix actually shipped, since a change can look complete in code without actually reaching the live product. Finally, we learned that small business owners value a tool that gives them one clear action, such as a specific reorder alert, far more than raw data or a dashboard full of numbers.

What's next for Stockpulse

Expanding item recognition accuracy across a wider range of shop types and product categories. Adding direct integration with real supplier ordering systems, so a reorder alert can turn into an actual placed order with one tap. Introducing sales pattern learning that improves depletion predictions the longer a shop uses the app. Building a simple multi location view for shop owners running more than one storefront.

Built With

• firebase-authentication
• gemini-api
• google-antigravity
• javascript
• postgresql
• rest-api
• supabase
• vercel