-
Recommendation Model
Inspiration
Most fashion platforms recommend what’s already popular. But what if you wanted to explore what’s not? Or discover fashion rooted in different cultures? We were inspired to create WeaverAI to flip the script—offering anti-recommendations, cultural fusion, and AI-driven fashion insights that embrace individuality and global expression.
What it does
WeaverAI is a smart fashion assistant that: Performs natural language search for fashion items using vector similarity. Generates anti-recommendations to help users explore unique, less popular fashion choices. Creates cultural fusion outfits by blending global fashion styles using AI Find your fashion twin—users with similar taste Provides choice approval ratings for your outfit or item decisions Gives actionable suggestions for fashion places and local stores
How we built it
**FastAPI** powers the backend APIs.
**Gemini AI (text-embedding-004)** is used for text-to-vector embeddings and prompt-based cultural fusion.
**Qdrant** handles high-speed vector similarity search (cosine distance).
**Qloo API** provides contextual insights and fashion places tagged with style categories.
We built a **Python CLI** tool for querying the system, useful during testing and demos.
Vector dimensions were reduced from 768 to 512 to fit Qdrant constraints while maintaining semantic meaning.
Challenges we ran into
**Embedding dimensionality:** Gemini returns 768-dim vectors, but Qdrant was optimised for 512-dim. We resolved this by vector truncation with normalisation.
**Semantic precision:** Generating meaningful cultural blends required prompt tuning to avoid vague or stereotypical outputs.
**Balancing API rate** limits across Qloo and Gemini, especially during batch requests.
**Anti-recommendation logic:** Flipping cosine similarity reliably while maintaining relevance was tricky and needed careful filtering.
Accomplishments that we're proud of
**Successfully integrated three advanced AI systems (Qloo, Qdrant, Gemini)** into a cohesive pipeline.
**Created a new paradigm in recommendation systems** by implementing anti-recommendations that are still user-relevant.
**Designed a seamless CLI interface** for interacting with the backend in real time.
**Developed an intelligent system** that goes beyond surface-level suggestions to uncover taste-based fashion intelligence.
What we learned
How to build and **fine-tune an AI-powered recommendation engine** using modern vector databases.
The importance of **prompt engineering and embedding manipulation for real-world use cases**.
How to **architect a backend that is modular, scalable, and future-proof**.
Real-world challenges of **dimensionality reduction, similarity scoring, and data filtering.**
What's next for WeaverAI
Building an **interactive web frontend with real-time visuals and outfit previews.**
**Integrating computer vision to provide instant feedback on user-uploaded photos.**
Enabling **user accounts, wardrobe histories, and personalised long-term insights**.
Extending **cultural blending with generative AI fashion design models.**
Hosting **fashion-based community interactions and curated trend discovery.**
Log in or sign up for Devpost to join the conversation.