Vlyzo

Your wardrobe, professionally styled for every occasion.

Comment

Inspiration

It's a common story: many guys especially in the tech world have money to spend but no idea how to actually dress well or up their game. Yet, so many people default to wearing the exact same black t-shirt every day or making random impulse buys that don't match anything they own.

Recognizing this massive market gap, we wanted to build a digital personal stylist. Vlyzo was built to stop the endless cycle of impulse buying and "nothing to wear" by helping people finally understand how to style the clothes they already own.

What it does

Vlyzo is an AI-powered styling assistant that acts as a pocket personal shopper. Users upload photos of their clothing, and the app automatically digitizes them into a sleek virtual wardrobe. The core magic is our AI recommendation engine, which mixes and matches your existing items to build fresh, confident outfits for any occasion or season. Going a step further, Vlyzo features an advanced Virtual Try-On experience allowing users to take a garment and generate a highly realistic image of it seamlessly draped on their own body, completely removing the guesswork from getting dressed or shopping.

How we built it

We built the Vlyzo mobile application with React Native/Expo, focusing on a premium, highly-polished user experience. The backend relies on Supabase for secure authentication, relational data, and edge storage.

For our heavy AI workflow, we built a custom FastAPI vision pipeline hosted on heavy GPU instances via Brev.dev. Here is how our custom AI stack breaks down:

  • Vision Pipeline: When a user uploads a photo, we run a multi-step sequence. We first use rembg (U2-Net) for background removal, followed by SegFormer B2 (trained on the ATR dataset) for pixel-precise clothing segmentation. Each segmented garment is then passed through FashionCLIP to extract deep semantic embeddings and classify attributes like category, style, color, pattern, material, and season.
  • LLM Styling Engine: For outfit recommendations, we use NVIDIA's Nemotron-Nano-9B-v2, served via vLLM. We feed the user's digitized wardrobe JSON and the specific occasion into Nemotron, taking advantage of its impressive context logic to intelligently generate cohesive, stylish outfit combinations.
  • Virtual Try-On: We integrated the state-of-the-art IDM-VTON pipeline on our Brev GPUs to take a garment image, a person image, and generate a flawless try-on composite natively within the app.

Challenges we ran into

Deploying a multi-model architecture was challenging. Orchestrating sequential AI steps (U2-Net → SegFormer → FashionCLIP) along with a dedicated vLLM server for Nemotron required careful resource management. Setting up the environment on Brev.dev to handle both transformers and Mamba-2 hybrid dependencies (for Nemotron) took significant debugging. Furthermore, managing the latency between the React Native frontend, Supabase edge functions, and the heavy GPU API required us to heavily optimize the payload sizes and base64 encoding logic to keep the mobile experience snappy.

Accomplishments that we're proud of

We are incredibly proud to have chained so many advanced, open-source AI models together into a cohesive, consumer-ready product. Going from a raw image upload directly to structured semantic data (FashionCLIP) and handing that context off to an LLM (Nemotron) to generate stylistic advice is a technical feat we're thrilled with. Integrating IDM-VTON to actually visualize those suggestions takes the app to the next level. We wrapped all this heavy computation in a stunning, glassmorphic UI that feels premium and effortless.

What we learned

We learned a tremendous amount about AI model orchestration, specifically how to manage GPU VRAM when loading multiple vision models alongside an LLM. We also gained deep practical knowledge of how to leverage Brev.dev for scalable AI deployment, and how to effectively bridge React Native frontends with hardware-intensive Python backends.

What's next for Vlyzo

Now that we have a platform that understands a user's unique style and existing wardrobe, the next step is connecting this personalized ecosystem with the broader fashion industry. We envision a future where fashion brands have a platform to understand what their target audiences are actually wearing. Using this data, brands can suggest highly specific "buying ideas" that perfectly complement the clothes a user already owns, helping the user further optimize their wardrobe while giving the brand an organic, hyper-targeted way to reach them.

Built With

Share this project:

Updates