Repair Before Replace: an AI-powered circularity assistant with persistent repair memory

This is a submission for Weekend Challenge: Earth Day Edition.

What I Built

I built Repair Before Replace, an AI-powered circularity assistant that helps people decide whether a damaged household item should be repaired at home, patched temporarily, or replaced responsibly.

The problem I wanted to solve is simple: people throw away useful items because they are unsure what is fixable, what is safe, and whether repair is worth it. I wanted to build something practical that nudges users toward repair first while still being honest about safety and limits.

The app lets a user:

• choose a supported category
• upload a photo of a damaged item
• get a structured damage assessment
• see whether the item is safe to repair at home
• get materials and step-by-step guidance when DIY repair is appropriate
• understand the waste impact of replacing versus repairing
• benefit from persistent memory across sessions

The differentiator is memory. The app does not behave like a stateless image classifier. It remembers prior attempts, preferences, and what worked before, then uses that history to influence future recommendations.

Examples of the memory behavior:

“Your previous glue-only patch failed after washing, so this recommendation uses hand stitching instead.”

“You’ve had success with wood glue and no power tools, so the same approach applies here.”

This is an AI repair companion that gets smarter the more you use it.

Demo

Live app: Repair Before Replace

Suggested demo flow:

1. Sign in
2. Select Furniture
3. Upload a photo of a damaged item such as a scratched table, cracked shelf, or broken chair
4. Review the assessment, repair guidance, and environmental impact summary
5. Scroll to the repair history and note how prior attempts influence the current recommendation

Home page with sofa loaded (Furniture selected, sofa photo, Analyze Damage button)

Code

GitHub repo: repair-before-replace

How I Built It

Product approach

I intentionally did not build a generic carbon calculator.

Instead, I focused on one real-world decision:

Should I fix this, patch it, get professional help, or replace it responsibly?

That led to a much more practical Earth Day project than a broad sustainability dashboard.

Front end

The UI is built with React 19 + TypeScript + Vite + Tailwind CSS v4.

I kept the flow simple:

• select category
• upload photo
• review structured assessment
• understand the next best action
• revisit repair history and memory-aware recommendations

A lot of the work here was product work, not just UI work: narrowing scope, separating safe DIY from professional repair, and making the recommendations feel trustworthy instead of overly confident.

Google Gemini

Gemini 2.5 Flash powers the core multimodal experience.

It analyzes the uploaded image and returns structured output for:

• visible damage
• confidence level
• safety check
• recommended action
• materials needed
• repair steps
• environmental impact band
• history-aware recommendation reasoning

One important design decision was using schema-constrained JSON output so the app could reliably render the same assessment structure every time.

Google AntiGravity IDE with Gemini AI

I used Google AntiGravity IDE with Gemini AI to accelerate the initial scaffold and part of the core implementation.

It helped compress the early build phases, including:

• app structure
• component setup
• routing
• service wiring
• rapid iteration on working UI flows

Gemini powers the runtime intelligence, while AntiGravity helped accelerate the path from concept to working prototype.

Backboard

Backboard powers the persistent memory layer.

Instead of storing repair history passively, I used it to make the next recommendation better.

Backboard stores things like:

• prior repair attempts
• what worked or failed
• user preferences
• category-specific history

Then that memory is injected into future assessments so the AI can reference earlier attempts directly.

That was a key product lesson for me: memory should influence the recommendation, not just create a history log.

Auth0 for Agents

Auth0 for Agents provides identity so the memory can stay personal and persistent across sessions.

Without identity, every user session becomes anonymous and the memory layer loses most of its value. Auth0 made it possible to keep repair history tied to a real user instead of treating every assessment like a first-time interaction.

Technical flow

User uploads photo
↓
Auth0 identifies the user
↓
Backboard fetches prior repair history for that user and category
↓
Gemini 2.5 Flash analyzes the image plus relevant memory
↓
App renders structured assessment, repair guidance, impact summary, and memory-aware explanation
↓
Completed assessment is written back to Backboard for future use

Stack

Runtime stack:
Gemini 2.5 Flash + Backboard SDK + Auth0 for Agents + React 19 + TypeScript + Vite + Tailwind CSS v4 + Firebase Hosting

Build tooling: Google AntiGravity IDE with Gemini AI for initial scaffolding and accelerated implementation

Scope choices

I kept the app intentionally narrow:

• Furniture
• Clothing & Textiles
• Cosmetic Appliance Damage

That let me focus on a believable UX and avoid overclaiming.

What I learned

The hardest part was not “getting AI output.”

The hardest part was building a tool that felt honest:

• separating repairability from safety
• making memory visibly useful
• avoiding fake precision
• narrowing the scope enough to keep trust high

That product discipline mattered more than any single model call.

Prize Categories

• Best use of Backboard — persistent cross-session repair memory that is injected into every AI call and visibly influences recommendations
• Best use of Auth0 for Agents — identity layer that keeps repair history personal and persistent across sessions
• Best use of Google Gemini — Gemini 2.5 Flash powers the multimodal image analysis and structured assessment flow, and Google AntiGravity IDE with Gemini AI accelerated the build process from scaffold to working prototype

Why I Built This

For Earth Day, I wanted to build something practical, not abstract.

A lot of sustainability conversations stay high-level. I wanted to make one real behavior easier:

fix more, toss less.

If a tool can help someone confidently save one chair, shelf, lamp, or household item from being thrown away, that already matters.

Youtube - Repair Before Replace