← Selected WorksGitHub

Computer Vision · 2025

Visual Search Engine

645ms visual jersey search at 8xSports — Letter of Recommendation.

Visual Search Engine — project screenshot
// Overview

Built during the 8xSports AI Engineering internship, this visual search engine locates jersey matches across 301 athlete profiles at 645ms end-to-end. YOLOv8 detects and crops jersey regions in a frame; DINOv2 extracts 384-dimensional feature embeddings; LBP texture analysis captures weave patterns; all vectors are indexed in FAISS for sub-millisecond approximate nearest-neighbour retrieval. The system runs in production at 8xSports and earned a Letter of Recommendation from the engineering team.

// Specs

Specifications

Employer8xSports (AI Engineering Intern)
Latency645ms end-to-end
Index Size301 athlete profiles
EmbeddingDINOv2 384-dim features
OutcomeLetter of Recommendation
// Features

Features

01

YOLOv8 object detection for precise jersey localisation and cropping

02

DINOv2 384-dimensional feature extraction for rich visual embeddings

03

LBP texture analysis for jersey weave pattern discrimination

04

FAISS approximate nearest-neighbour index across 301 athlete profiles

05

645ms end-to-end production latency under real workload

06

Earned Letter of Recommendation from 8xSports engineering team

// Tech

Tech Stack

YOLOv8DINOv2FAISSOpenCV

Interested in
Visual Search Engine?

Explore the source code, architecture notes, and full implementation.

View on GitHubGet in Touch →