Inspiration
Getting a good deal on something you've always wanted (used) has always been hard. You end up cross-checking prices across platforms, comparing them with retail, squinting at blurry photos for wear, and trying to avoid the shadiest deals. Detective Sleuth does all of this for you as you browse. It's an accelerated browsing companion, and a sceptical second opinion before you buy.
For those on tight budgets, or wanting to save money where buying new can't really be justified!
What it does
- A browser extension for eBay and Vinted listings and search results
- Set your size, colour, budget and priorities once; they stay on your device
- On a search page, it narrows the results down to 8 worth looking at
- On a listing, it investigates:
- Spots defects and boxes them on the seller's photos
- Identifies the model from the listing, using Google Lens when it's vague
- Gemini Flash and Pro cross-check the model, and it asks you to confirm when unsure
- Checks your size and what reviewers say about comfort
- Compares the price with the same shoe new, in the same colour
- Flags warning signs, like suspiciously low prices, without calling anything fake
- Gives an evidence-backed verdict and suggests questions for the seller
How we built it
Frontend: Chrome extension side panel (TypeScript, React, esbuild)
Backend: Node.js and TypeScript on Next.js route handlers, with Zod
AI: Gemini Flash and Pro
Data: SerpApi (Google Shopping, web search, Google Lens)
Database: PostgreSQL on Neon via Prisma
Challenges we ran into
- Navigating legal issues surrounding using a remote scraper from eBay/Vinted, and storing that information on our servers. I moved away from remote scraping completely: Sleuth only reads the page you're on when you ask it to, and strips out seller details before anything is stored or sent to the AI.
- Gemini can be confidently wrong. One model named the wrong version of a shoe with 90% confidence, which is why we added the cross-check between Flash and Pro, and why Sleuth admits when it isn't sure
- Our first full investigation took over a minute. I got it down to around 20 seconds by merging steps and running checks in parallel.
Accomplishments that we're proud of
- Making it useful
- It works on real, live listings on eBay and Vinted, not just a demo
- Every verdict comes with its evidence, and when Sleuth can't judge something, it says so instead of guessing
- It's built with fairness and privacy in mind: seller details are removed, your profile and history stay on your device, and it never accuses anyone (including actual scammers)
What we learned
- How to navigate data responsibly
- AI's own confidence score can't be trusted, it has to be cross checked with a more solid, deterministic rule base or another AI.
- To let AI handle the messy parts and let plain code handle anything that has to be exact, like sizes and prices
What's next for Detective Sleuth
- Adding categories beyond trainers, like clothing and tech
- More marketplaces, such as Depop and Facebook Marketplace
- Letting user corrections improve Sleuth over time, once they've been reviewed
- Comparing against what items actually sold for, not just what sellers are asking
Built With
- chrome
- esbuild
- gemini-api
- neon
- next.js
- node.js
- postgresql
- prisma
- react
- serpapi
- typescript
- vitest
- zod