Projects

Exquisitor

Exquisitor is an interactive visual search and exploration system I co-created for navigating large-scale image and video collections — often containing millions of items — in real time.

The system combines two complementary interaction paradigms:

• Conversational search — users describe what they're looking for in natural language, and the system retrieves semantically relevant results using large vision-language models.
• Relevance feedback — users refine results by marking items as relevant or not, allowing the system to progressively home in on what they actually want.

Together, these make Exquisitor effective for both precise known-item search and open-ended exploration, where users may not have a fully-formed query at the start.

Competition results

Exquisitor has competed in several international multimedia benchmarking competitions, including the Video Browser Showdown (VBS) and the Lifelog Search Challenge (LSC), consistently ranking among the top systems.

How it works

At its core, Exquisitor uses a hierarchical indexing structure that enables sub-second retrieval over millions of items on commodity hardware. Conversational queries are encoded with foundation models (e.g., CLIP) and combined with relevance feedback signals to iteratively rerank the collection — bringing the most relevant content to the surface with each round of interaction.

The backend is architected as a SOLID-compliant FastAPI server, with extensive documentation available at docs.exquisitor.org.

Visit exquisitor.org →