Computer Vision Engineer — Visual Search & Discovery

Hugo Lindqvist

Contract · Mid-level · New_York

This role is no longer accepting applications

Similar open roles you can apply to now:

About the role

We are a fashion discovery platform. We have 6.2 million registered users who use our app primarily to find products visually — photographing something they like in the real world and finding shoppable equivalents from 3,400 brand partners. Visual search is the product. The accuracy and speed of our image matching and similarity ranking directly determines whether users come back. Our current visual search pipeline is built on a contrastive learning model we trained eighteen months ago. It performs well on our top 40 brand partners whose catalogue images are high-resolution, consistent, and well-structured. It performs poorly on our long-tail brand partners whose catalogue imagery is inconsistent in background, lighting, and angle. We are hiring a contract Computer Vision Engineer for a 10-month engagement to improve similarity search performance specifically on long-tail catalogue imagery, and to build the evaluation infrastructure to measure improvements systematically rather than anecdotally. The work involves both model improvement and the unglamorous but essential work of building a test set that actually measures what the users experience.

Responsibilities

Improve visual similarity performance on long-tail brand catalogue imagery — fine-tuning, augmentation strategy, and training data curation
Build a systematic evaluation framework for visual search quality using real user query logs and human relevance judgements
Optimise approximate nearest neighbour search for new embedding space — index construction, quantisation, and recall/latency trade-offs
Work with the permanent engineering team to integrate model improvements into the live search pipeline
Deliver documented model architecture, evaluation methodology, and operational runbook at contract end

Requirements

4+ years of computer vision engineering with experience in image similarity, visual search, or contrastive representation learning
PyTorch for fine-tuning and evaluating vision transformers or contrastive models on fashion or product imagery
Python for evaluation pipeline development, data augmentation tooling, and model serving integration
FAISS or equivalent for approximate nearest neighbour search at catalogue scale
Scikit-learn for baseline evaluation metric calculation and statistical comparison of model variants
Experience building image evaluation datasets and automated visual similarity benchmarks — this is as important as the modelling work
Documentation discipline appropriate to a contract engagement — the permanent team must be able to extend what you build