Velo | Devpost

Inspiration

We want to created an all in one shopping platform to solve a lot of the annoyances involved with modern day online shopping. Things like needing 10+ tabs across stores to piece together one outfit, recommendations that ignore your taste and budget and no way to see it on you, or with the rest of your look.

What it does

Velo is a mobile app that pulls clothing from major retailers into a single feed built around you. You start with a short questionnaire covering style (skate, rock, etc.), brands you like, budget, and colors, and you pick from a set of style examples. Then you upload one full-body photo.

From there, Velo builds your home feed as a grid of items. On any item you can:

Save it to your cart
Discard it so it stops showing up
Hit "more like this" to pull a new batch based on that piece
Type what you want changed in the search bar ("less formal," "under $40," "more earth tones") and the feed re-ranks

You can also see the item rendered on your own photo through virtual try-on. When you're ready to buy, Velo sends you straight to the retailer's product page.

How we built it

We made Velo into three services:

Phone app (Expo / React Native): the style quiz, photo upload, feed, showcase and cart.
Backend (FastAPI + SQLite): stores everything and finds real products on Google Shopping. It's the only part the app talks to.
AI service (FastAPI + OpenAI): plans the searches, ranks products for each user, builds outfits and renders try-on images of the user.

Challenges we ran into

We found a lot of problems with latency. Planning, searching, and ranking a full feed isn't instant, and try-on renders are slower.
Translating taste into searches. Getting from "I like skate style, earth tones, under $50" to Google Shopping queries that return good results took a lot of iteration on the planning step.
Keeping the architecture clean under hackathon pressure. Routing everything through the backend meant writing more endpoints, but it kept the AI service and database locked down.

Accomplishments that we're proud of

Every way a user can steer the feed with our recommendations engine mangaging to make really specfic and detailed suggestions to users.
Not only can users see real retail items rendered on our own photos but they with the recommendations outfits can be auto selected and placed on the user
Going from a rough list of ideas to a mapped-out API before writing feature code, which saved us a lot of rework.

What we learned

Designing the API up front, down to which service is allowed to talk to which, made splitting work across the team much easier.
LLMs are much more reliable at turning fuzzy feedback into structured parameters than at doing the whole job in one shot. Breaking recommendation into plan and rank steps made each piece easier to debug.
Anything slow belongs in a background job. Users will wait on a loading screen, but they won't wait on a frozen one.