Inspiration

Everyone knows a Frank. His chair wobbles, so he has two options: spend $289 on a new one, or watch a 47-minute tutorial and somehow end up with a leg above the backrest.

Most broken things aren't really broken. A chair wobbles because one screw backed out. A faucet drips because a 30-cent washer wore out. But most people can't tell what is wrong, so they replace the whole thing. That's expensive, and it's a lot of furniture in landfills.

I wanted the shortest possible path from "this is broken" to "this is fixed": just film it.

What it does

  1. Film the broken thing. Upload a short video or photo, plus an optional note.
  2. Get a diagnosis pinned to your video. FixFrame finds each fault and boxes it at the exact second it shows up, like "screw backing out of the joint at 0:05". You also get a confidence score, the tools you need and a safety note.
  3. See repair vs. replace. For Frank's chair: about $9 to repair vs. $289 for a new chair. That's $280 saved and 6.5 kg kept out of landfill.
  4. Follow step-by-step instructions. Click any step and the video jumps to the moment that shows it.
  5. Buy the exact parts. Channel3 searches real stores for each part. You compare offers in an AG Grid table, and you can even search by a photo crop taken from your own video.
  6. Pay with PayPal. One checkout for all the parts.
  7. Or hire a pro, and pay only if it's fixed. PayPal authorizes the pro's quote but doesn't charge it. After the repair, you film the item again. FixFrame compares the after-clip to the original problem:

$$ \text{payment} = \begin{cases} \text{capture (pro gets paid)} & \text{if fixed with confidence} \geq \tau \ \text{void (you pay nothing)} & \text{otherwise} \end{cases} $$

No more paying for a "fix" that falls apart the moment you sit down.

How I built it

  • Backend: Node.js + Express. Uploaded videos go through ffmpeg, which pulls evenly spaced frames with timestamps and crops regions for photo search.
  • Diagnosis: the frames go to a multimodal AI model (Anthropic API) with a strict JSON schema: item, problem, markers with timestamps and bounding boxes, steps, parts, tools, safety notes and a replacement estimate. The server validates and cleans the response before the UI ever sees it.
  • Parts: each part becomes a Channel3 search, by text query or by a photo crop from the frame where the fault appears. The cheapest in-stock offer is picked by default.
  • Comparing offers: AG Grid with custom cell renderers for the pick buttons, product thumbnails and stock labels. Change a pick and the cart total updates live.
  • Payments: PayPal Orders v2. intent: CAPTURE for parts, and intent: AUTHORIZE → capture or void for the pay-only-if-fixed flow.
  • Hosting: Render, with a Docker blueprint so ffmpeg is available in production.
  • Frontend: plain HTML, CSS and JS, with an editorial design, and a video player with a marker timeline that stays in sync with the step list.

Savings are calculated as:

$$ \text{saved} = P_{\text{replace}} - \sum_{i} \min_j \, p_{ij} $$

where $p_{ij}$ is the price of offer $j$ for part $i$.

Challenges

  • Pinning faults to the right moment. Getting the model to say what's wrong was easy. Getting it to say when and where in the video was hard. I send the timestamp with every frame and ask for normalized bounding boxes, then clamp and validate everything.
  • Turning a diagnosis into something you can buy. "Loose joint" isn't a product. Each part needed a shopping query specific enough to return wood glue and #8 screws, not a whole new chair.
  • Getting pay-only-if-fixed right. Authorize, then capture or void, sounds simple. But the timing of each step, and showing the user clearly what's held vs. charged, took a few rewrites.
  • Making the demo reliable. Live AI and live search can be slow or fail during judging, so FixFrame recognizes its sample clips and has a fallback mode. The app always works, even without keys.

What I learned

  • Using PayPal's authorize-and-capture flow as a trust tool, not just a way to take payments.
  • Prompting a vision model for structured output grounded in time and space, and never trusting that output without validation.
  • That the best demo is a story. Nobody gets excited about "multimodal diagnosis", but everyone gets Frank.

What's next

  • Live video mode: point your phone at the problem and get guided through the fix in real time.
  • A repair history for each item, so FixFrame knows the chair you fixed last spring.
  • A pro marketplace with verified reviews based on how many repairs actually passed the after-clip check.

Built With

Share this project:

Updates

Submission history