RepairVision: An Intelligent Repair Café Platform for Multimodal Fault Diagnosis and Sustainable Electronics Repair

Every week, perfectly fixable things end up in the bin. A kettle with a tripped fuse. A lamp with a loose wire. A jumper with a hole in the elbow. Repair cafés exist to stop that: free, volunteer-run sessions where people bring something broken and fix it together.

For Hacktoberfest Hack Day Coimbatore 2026, our team, The Blacklisted, built RepairVision: a platform that runs a repair café day to day and adds AI where it genuinely helps: working out what is wrong, and deciding whether a repair is worth it.

The problem

Repair cafés do great work, but three things hold them back:

Running a session is chaotic. Paper sign-in sheets, a queue of visitors, keeping track of who is fixing what.
Diagnosis depends on who turns up. The volunteer who knows bikes might not know hairdryers.
People don't know if repair is worth it. Without a clear comparison, the default is "just buy a new one".

What we built

RepairVision follows an item from "it's broken" to "it's fixed":

AI Diagnosis

Describe the device and add a photo. Gemma 4 suggests up to three likely faults, each with the evidence that supports it and the evidence that's still missing, then recommends safe checks and asks one question at a time to narrow it down.

Every result is labelled as provisional, and a hazard screen that doesn't depend on the model keeps risky devices (mains appliances, e-bike batteries) on the safe side. Each suggestion is backed by our own repair knowledge base of 100+ entries, the café's past repairs and iFixit guides.

Repair or Replace?

The question people actually ask next: is it worth fixing? The advisor compares the repair cost with a replacement, allows for the chance the repair fails, and shows the CO₂ and waste a repair avoids, with all the working shown.

A site assistant that knows what's happening

A friendly 3D character sits in the corner of every page. Ask it "What events are on right now?" and it answers from the café's live data, then opens the right page for you.

It only ever sees what the public website already shows, so it can't leak anyone's name, contact details or repair, whatever you ask it.

Everything a café needs on the day

QR check-in, a live repair queue, a waiting-room display, session planning and impact statistics. All of it is in one place, in a clean dark interface.

How we built it

Frontend: SvelteKit, TypeScript and Tailwind CSS
Backend: a single Cloudflare Worker serving the pages and the API
Data: Cloudflare D1 (SQLite) with Drizzle ORM, and R2 for photos
AI: Google Gemma 4 (gemma-4-26b-a4b-it) through the Gemini API

A few decisions we're proud of:

We never trust raw AI output. Gemma's replies must be valid JSON that passes a schema, and safety rules run in our own code afterwards.
No invented answers. If the AI is unavailable, people see an honest error. Prices are never made up.
Privacy by design. The public assistant reads only public data, and photos sent for diagnosis aren't stored.
Free to run. Cloudflare's free tier plus Gemma via the Gemini API means a volunteer café can host it at no cost.
Tested offline. All 199 tests use fake AI responses and run in the real Workers runtime.

Challenges and what we learned

Making AI safe to show to the public took more than a good prompt. Strict schemas and code-level guardrails made the difference.
Four people, one codebase, one day. Small focused commits, shared types in one package and a fast test suite let us merge each other's work without breaking things.
Free-tier limits shape design. Working within about 10 ms of CPU per request pushed us towards caching and keeping AI calls as network waits rather than computation.