Muse.AI

Inspiration

In 2025, my friends and I decided to go to Coachella together with everyone matching a specific theme for each day of Coachella. These themes varied from cowboy, to apocalyptic, to Y2K. Most of us found it challenging to imagine what the theme is and even more so, where to get the items. Many problems persisted with sizing and availability. Whole process was a bit of a mess. We all somehow managed to get the outfits needed for each day, but was not the most pleasant experience. This sprung the idea of how to more conveniently source outfits and choose the ones that fit your style more without all the hassle of sourcing each item one by one.

What it does

Muse is a fashion stylist application curating outfits or items based on a desired theme. It outputs the image of the look along with direct links to shop the specific items.

How we built it

I built Muse as a system of specialized AI agents that work together like a relay team. To keep them on the same page, I created a custom Style Guide that defines 16 specific fashion aesthetics (like Boho or Streetwear) so the AI understands style exactly like a human stylist would.

• The Eyes (Visual Analysis) This agent looks at your photo to figure out your general vibe and size. If you upload inspiration pictures, it understands that you might like a mix of styles—combining them rather than forcing you into just one category.

• The stylist Synthecize outfit recommendations based on user's information such as occasion, style, gender, etc.

• The Shopper (Procurement) This step splits the work. Separate agents go searching for each item (shoes, tops, etc.) using Google Search. They look for exact matches first, but if those are sold out, they find the next best available option that you can actually buy.

• The (Curation) This agent combines speed with taste. It instantly filters out broken links or items that are over budget, then uses a smarter AI model to look at the full outfit and give it a "visual score" from 0-100 based on how well the pieces match.

• The Manager (The Director) This is the final quality check. It reviews the budget, color coordination, and links one last time to make sure the outfit makes sense before showing it to you.

• The Engine (Infrastructure) The app is built to be fast. It saves recent searches so it doesn't have to do the same work twice and uses modern web technology to handle the heavy lifting of searching through thousands of items instantly.

Challenges we ran into

• Structured Output The biggest challenge was getting the AI to return structured, usable output instead of free-form text — I spent more time engineering prompts and building fallback parsing than writing actual features.

• Fashion Data There's also no clean "fashion API" that hands you shoppable items, so I had to stitch together Google Search grounding and filter out junk sites manually.

• Agent Chaining And when you chain multiple AI agents together, one bad output cascades through the entire pipeline, which forced me to build verification layers between every stage to catch errors before they snowballed.

Accomplishments that we're proud of

What we learned

25, my friends and I decided to go to Coachella together with everyone matching a specific theme for each day of Coachella. These themes varied from cowboy, to apocalyptic, to Y2K. Most of us found it challenging to imagine what the theme is and even more so, where to get the items. Many problems persisted with sizing and availability. Whole process was a bit of a mess. We all somehow managed to get the outfits needed for each day, but was not the most pleasant experience. This sprung the idea of how to more conveniently source outfits and choose the ones that fit your style more without all the hassle of sourcing each item one by one.

What I Learned

I started this project with zero coding experience, so the learning curve was quite steep. I familiarized myself with React and TypeScript just to get the visuals working, then figured out how to connect the APIs to make the system function.

• Prompting   Working with Gemini showed me that an open-ended question is not very valuable. I had to engineer specific formats and rules to force the model to give me reliable data I could actually use in the code.

• Data   I discovered there is no easy process for product data. I had to piece together Google Search results myself and write logic to handle the messy reality of broken links and out-of-stock items.

• Agents   The multi-agent setup forced me to break complex goals into small steps. I learned to split tasks so they could run at the same time, which kept the app from freezing while it thought.

• Speed and Cost   I learned that saving results (caching) is valuable. Without it, every minor tweak I made forced the AI to redo all its work, burning through money and time.

What's next for Muse.AI

Will add a gallery or ranking system where users can share and see other people's fashion ideas and like them. By doing so, we create a community and the app will learn users' tastes better to provide better recommendations.

Built With

• google-gemini-api
• google-programmable-search-api
• react
• tailwind-css
• typescript
• vite