FitCheck 讓用戶出發取貨前在房間預覽二手傢俱
FitCheck: see secondhand furniture in your room before you drive to get it
FitCheck 整合 Craigslist、Facebook Marketplace 和 eBay 的二手傢俱搜尋,讓用戶以 AR 按實際尺寸在房間預覽物品,並查看運送或自駕成本及車廂是否放得下。項目展示時收集了 1,070 個獨特刊登,搜尋中有 21 個 3D 模型,典型搜尋約需 0.7 秒。
This is a submission for the MLH x DEV Writing Challenge
What I Built
FitCheck lets you see secondhand furniture from Craigslist, Facebook Marketplace, and eBay in your own room, at real size, before you drive across town for it.

There are two bottles on this table. Only one of them exists. The other is a 3D model FitCheck built from one photo, placed in AR at its real size.
The idea came from move-out week. Every June the curbs around campus fill up with desks, lamps, and nightstands, while a mile away someone orders the same nightstand new. The EPA puts US furniture waste at about 9.7 million tons a year, and most of it ends up in landfill. When we tried buying used ourselves, the same three things kept stopping us:
- Listings are spread across three sites, so you run every search three times.
- A photo can't tell you whether a dresser fits next to your bed, or in the back of your Civic.
- The listed price isn't what you pay. A $20 table 25 miles away costs more once you count the drive, and half the time it's already gone when you message the seller.
So here's what FitCheck does.
Search by voice. Hold the mic and say "desk lamp and side table under forty dollars, warm wood." The filters show up as chips you can edit, and results from all three marketplaces come back in one list.
Show the real cost. Every card shows the price plus the round-trip drive (or shipping, if it ships), and a badge for whether it fits in your car with the seats up, with them folded, or not at all.
AR at real size. Tap "See it in your room," and the item opens in AR Quick Look on iPhone or Scene Viewer on Android, with no app to install. The model is built from that listing's own photo and scaled to its measured width, depth, and height.
Check it's still listed. Opening a listing fetches the original page live, so you don't drive to something that sold yesterday.
Sell. Take one photo of anything, a chair or a water bottle. Gemini estimates the dimensions, FitCheck suggests a price from comparable local listings, and a 3D model gets built while a progress bar counts down. One tap copies the listing text and photo for Marketplace, Craigslist, and eBay.
By demo time, we had 1,070 unique listings pulled from about 70 scrape jobs, every one tagged and searchable by meaning, and 21 real 3D models live in search. A typical search came back in about 0.7 seconds.
The biggest thing I learned is that most of the work in an AI project happens around the model. Checking Gemini's dimensions against sane ranges, marking estimates as estimates, refusing to show a generic chair in AR when we didn't have a model of the real one: those decisions took longer than any API integration, and they're why the demo felt trustworthy. I also learned to keep slow work out of the request path. A 3D model takes a couple of minutes to build, so the website writes a job row to Postgres and a Python worker picks it up. The site never sits and waits.
Demo
- Live site: fitcheck-sd.vercel.app (open it on your phone to try the AR)
- Code: github.com/abhishaiganta/FitCheck
- Devpost: devpost.com/software/fitcheck-1jfawd

The Sell flow, end to end: one phone photo in, and a couple of minutes later a true-to-size model stands next to the original. Same rope handle, same D-ring.
How the pieces connect:
Bright Data (FB Marketplace, eBay, Craigslist)
│ raw listings
▼
Python worker ── Gemini vision tags + embeddings ── cutout → image-to-3D → GLB + USDZ
│ │
▼ ▼
Neon Postgres + pgvector (listings, vectors, job queue) Vercel Blob (photos, 3D files)
│ │
▼ ▼
Next.js on Vercel: ElevenLabs STT → Mastra agent on Gemini → one hybrid SQL query → ElevenLabs TTS
│
▼
Phone: <model-viewer> → AR Quick Look (iOS) / Scene Viewer (Android)
Partner Technologies
Google Gemini
Gemini does four jobs in FitCheck.
1. Tagging every listing. Marketplace data is messy. 84% of our listings had no dimensions at all, titles say things like "OBO", and the photos are cluttered. Our Python worker sends up to three photos per listing (resized to 768 px) plus the title and description to a Gemini vision model. It returns category, style tags, condition, estimated weight, width, depth and height, the best photo, and whether the item is a good candidate for 3D. We clamp the numbers to sane ranges and store where each dimension came from, the seller's own text or Gemini's estimate, so the app can label it.
2. Search by meaning. gemini-embedding-001 turns every listing into a 768-dimension vector stored in Neon with pgvector. When a search includes a style like "warm wood", we embed that phrase and rank by cosine similarity inside the same SQL query that applies the budget, distance and car-fit filters. That's how "warm wood" finds a walnut nightstand whose title never mentions wood.
3. Understanding the request. A Mastra agent running on gemini-flash-lite-latest turns the spoken sentence into a typed tool call: categories from a fixed list, a budget, a style phrase, a car. The tool schema is strict (Zod), so the model can't invent a category or hand us a malformed price, and it never writes SQL. Our code runs the query. If the agent fails, search falls back to a single direct Gemini function call, then to a plain keyword parser, so it keeps working either way.
4. The Sell flow. Gemini measures the item in the seller's photo. When there aren't at least three comparable local listings to price from (nobody else is selling a water bottle), it estimates a used price instead, and the page tells the seller which method it used.
Working with it: flash-lite is fast enough that it barely shows up in our search time. What did catch us out was quota. Free-tier limits apply per model and per project, so both the worker and the web app ended up tracking usage per model, rotating across keys, and backing off with cooldowns rather than retrying into a 429. Next time I'd build that bookkeeping on day one, not at hour twenty.
ElevenLabs
Voice is the front door of the app. The browser records with MediaRecorder and posts the audio to our own /api/stt route, so the API key never reaches the phone. The route checks the length and format, then sends it to ElevenLabs speech-to-text (scribe_v1). The transcript goes straight into the Gemini pipeline above.
When results come back, ElevenLabs text-to-speech (eleven_flash_v2_5) reads a one-line reply: "Found 3 desk lamps and 2 side tables under $40." Our first version also read out the cheapest price and the distance. In testing that was too much to sit through, and it repeated what was already on the cards, so we cut it down to the count. Shorter replies also make the flash model feel instant.
Working with it: the API itself was the easy part. The gotcha was audio formats. Safari on iPhone records mp4/m4a and Chrome records webm, so the route maps MIME types explicitly instead of trusting whatever the browser says.
Hackathon Experience
We built FitCheck at SanD Hacks in San Diego as a team of four: Avanthika Daita, Avyay Bharath, Varun Ponnala, and me.
Attending SanD Hacks was such a wonderful experience, as it gave me the opportunity to meet with industry professionals as well as work together in a fast-paced setting to design and build a product that solves a real-life problem.
The moment I'll remember is the first time a model from a real Craigslist listing showed up on our floor at the right size. Until then it was all logs and database rows. After that it was a lamp you could walk around. Close behind: our whole 3D pipeline ran on one laptop that was not allowed to sleep for the entire weekend, and we spent part of the night figuring out how to mirror an iPhone into Zoom so the judges could see the AR.
We ended up winning the Data for Good track, sponsored by Bright Data.
Bright Data's Datasets API pulled our Facebook Marketplace and eBay listings, and its Web Unlocker got us into Craigslist, which blocks ordinary scrapers. The Web Unlocker also powers the live "still listed?" check.
If you're furnishing a place on a budget, try it on your phone at fitcheck-sd.vercel.app, and tell me what breaks.
來源:DEV Community · dev.to
