Mendly

Point at what’s broken and get the fix, step by step, plus the right replacement part at the best price.

Comment

Inspiration

My dad and I buy cars at auction and fix them up together. Every time something breaks, we go through the same ritual: figure out what's wrong, find the right part, and make sure it actually fits. The part that fits is always the hard bit. One wrong part and you've wasted money and a weekend.

Then I noticed it's not just cars. People throw out faucets, microwaves, and fridges every day because they have no idea the fix is easy and cheap. So I built Mendly: point at what's broken and get the fix, step by step, plus the right replacement part at the best price.

What it does

  1. Snap a photo of the broken thing and say what's wrong in a few words.
  2. Mendly figures out what the item is and which part has probably failed.
  3. It asks for your brand and model number, because a part that fits one model can be useless for another.
  4. You get a safety-first repair plan with tools and tappable steps to check off as you go.
  5. It finds the exact part for your model, with a price and a link to buy it.
  6. It tells you whether to fix or replace, by comparing the part's price with a brand-new one.

The fix-or-replace call is simple math. If the part costs

$$\frac{\text{part price}}{\text{new price}} \le 0.4$$

then Mendly says Fix it. Between 0.4 and 0.75 it's a close call, and above that it says to consider replacing.

How I built it

Mendly is a mobile-friendly website, so it works on any phone or computer with nothing to download. Everything runs on Nebius Token Factory using three NVIDIA Nemotron models, each picked for a different job:

  • Nemotron 3.5 Super VL looks at the photo.
  • Nemotron 3 Super 120B reasons about the repair and writes the plan.
  • Nemotron 3 Nano quickly pulls parts and prices out of search results.

Tavily searches the live web at runtime. It first looks for the parts list for your exact model, then reads store pages to find prices. The backend is Python with FastAPI, and it's hosted on Render. A daily usage limit keeps the free demo from running out of credits.

Challenges

  • iPhone photos are secretly HEIC. A photo named .jpg can break everything. Mendly now converts any photo it receives.
  • AI sometimes returns messy JSON. I added retries and a fallback to the bigger model so the app doesn't crash.
  • Generic results are the enemy. Asking for a "fridge light bulb" returns ordinary bulbs. Fixing that meant searching the model's own parts list, dropping unrelated results, and only marking a part "Matches your model" when a page actually mentions the model number.
  • Model numbers are messy. Labels like LFDS22520S /04 have suffixes that parts sites leave off, so Mendly tries shorter versions of the number too.
  • Faucets live on different sites than appliances. I added plumbing parts sites, and Mendly now also looks up the product line, like Moen "Arbor".
  • No made-up prices. Mendly only shows a part, price, or link that came from a real search result. If it can't find a price, it says so.

What I learned

Getting the right part is a harder problem than getting a good answer. The AI part was the easy bit. The real work was making Mendly honest about what it doesn't know, like asking for a model number and saying "check fit" when it isn't sure.

What's next

Searching manufacturers' own parts sites, a "stuck on a step?" chat with photos, and support for more categories, from bikes to car electronics.

Built With

Share this project:

Updates

Submission history