How we built Fynd Now using Visual Similarity

At Fynd we build technologies to help customers discover fresh fashion. Our marketplace channel Fynd
By Debajit Sardar
January 9, 2020
6 min read

At Fynd we build technologies to help customers discover fresh fashion. Our marketplace channel Fynd - Online Shopping App went live in Oct 2015 on playstore and appstore followed by an early 2016 release of the web version fynd.com. After that the team has put significant efforts in creating innovative features for customers such as Mix & Match, Gravity View, Shop by Categories, The Closet etc. The goal has always been to make the discovery easier for customers. A recent addition to Fynd’s feature list was Machine Learning backed recommendation engine which provides personalised product feed based on customer behaviour.

A special Applied Machine Learning (AML) engineering domain was already working on various fields of ML/AI over the years. With persistent effort in research and development the AML team has published 5 research papers and built numerous ML models. The team has added visual similarity features on Fynd Marketplace aswell which has improved contextual product recommendations. Give a visit to research.fynd.com to view projects, tools, blogs that the AML team has built so far.

Last few months designers and AML engineers are working closely to leverage the visual similarity ML model for customers to easily shop similar products from inspirations they draw on the internet. The model was already trained with huge datasets of fashion photos and e-commerce product images. All we needed was a product to serve the customers.

How we built itWe wanted to build a standalone product which should be effortlessly accessible to customers while browsing on the web. At work we use Google Chrome extensions such as pinterest, grammarly etc. A simple chrome extension was our solution too.

The challenge for engineers was to enrich the ML models for the highest precision and performance. FyndNow should  process any fashion image in real time and fetch similar products from 1 lakh + products in Fynd catalogue. Engineers fine tuned the ML filters such as colours, patterns, gender, neck type, collar, sleeve lengths etc. to accurately generate relevant search results.

Meanwhile designers started making prototypes for a  discovery experience to browse similar products on top of the website which the customers is currently active on. The experience needed to be uninterrupted. That’s the reason major variations in between the prototypes were based on how much real estate the product should occupy. Most importantly the success of the product will be measured based on the product views it generates on Fynd Marketplace.

Fynd Now extension is now live on chrome store and is publicly available on now.fynd.com.