← Back to portfolio

Project case study

LookR

Fashion search with computer vision and generative AI.

View source code ↗

01

Problem

Finding visually similar clothing or expressing a visual idea through text is hard with traditional filters.

02

Solution

LookR combines text, image, and inspiration search with visual and semantic embeddings to recommend relevant garments.

03 · Architecture

Image/Text input -> FastAPI -> CLIP + semantic embeddings -> Pinecone -> Ollama recommendations

04

My contribution

I designed part of the application architecture, evaluated different LLMs and VLMs, refactored backend routes, developed the agent and its prompt, and made configuration and integration improvements to Pinecone.

05 · Stack

PythonFastAPICLIPPineconeOllamaEmbeddings

06

Technical decisions

Visual and semantic embeddings remain separate to combine complementary signals and enable image, text, or mixed queries.

07

How it works

Text, image, or inspiration input is represented and queried against Pinecone indexes. The service returns relevant garments while the AI agent adds context; each wardrobe stays isolated in a private namespace.

08

Technical challenges

  • Evaluating LLMs and VLMs for quality, latency, and local resource use.
  • Supporting visual, semantic, and mixed queries.
  • Keeping private Pinecone namespaces isolated.

09

What I would improve

  • Reproducible relevance evaluation.
  • Authentication and persistent favorites.
  • Independent deployment of inference services.