I am Gregory Le Goff, a Data and Machine Learning engineer and the founder of RecoKit.
Throughout my career, I have worked on the same underlying challenge in different forms: how can an algorithm make a sound decision when the available data is insufficient?
Sales forecasting, Supply Chain optimization, recommendation engines, new products and the long tail have kept bringing me back to this question. It is also what led me to create RecoKit.
From Data to AI applied to Retail
I began my career at Sopra in 2002, working mainly on decision-support architectures, Data Warehouses and large-scale data processing.
In 2010, I joined Vekia, then a young company applying Machine Learning to Retail and Supply Chain problems. For seven years, I worked on sales forecasting and Supply Chain optimization as the field evolved from statistical models to SVMs, Random Forests and Deep Learning.
This experience shaped how I approach AI: start with a concrete business problem, build the mathematical model that can answer it, and make it work reliably in production.
Recommendation algorithms at scale
After working on Big Data architectures at Oney, I joined the Leroy Merlin / ADEO ecosystem in 2018 and specialized in large-scale product recommendation systems.
Collaborative filtering, embeddings, graphs, Computer Vision and language models evolved over time. One difficulty remained unchanged.
How do you recommend a product that has just arrived?
A traditional recommendation engine performs well once it has enough clicks, orders and user behavior. A newly listed product has none of that history. This is the cold-start problem.
For an e-commerce SMB, the issue may affect not just one product but a large part of the catalog.
Why I created RecoKit
Large retailers can compensate with billions of behavioral events. A store with a few hundred or thousand products cannot.
RecoKit takes a different path. Instead of waiting for customers to generate enough data, it first understands the products themselves: their purpose, uses, attributes and relationships.
Language models, embeddings and behavioral evidence can then be combined progressively. This lets a merchant obtain relevant recommendations from day one and improve them automatically as real store data accumulates—without reducing discovery to existing best-sellers.
Research, production and teaching
Alongside my industry work, I have taught at the University of Lille since 2004, including data science, computer science and the methods required to turn an algorithmic idea into a usable system.
I do not see AI as a succession of fashionable technologies. I care about understanding a problem, finding the right mathematical or algorithmic representation, and building a system simple and robust enough to work in practice. RecoKit was born from this approach.