PyData Amsterdam 2024

Retrieve me if you can: SLM-powered retrieval to scale freelancers matching at Malt
09-20, 11:50–12:25 (Europe/Amsterdam), Rembrandt

This talk unveils Malt's secret weapon for highly efficient freelancer matching - a powerful neural retriever. We'll showcase how Small Language Models (SLMs) help us instantly connect companies with their ideal freelancers. Discover how our retriever was built, deployed with a vector database and optimized to minimize resource consumption.


Ranking all items in a large-scale recommender system is impractical. To tackle this challenge, a common approach is to employ a candidate generation phase, where only the top candidates are ranked. In this talk, we’ll leave aside the ranking phase to explore the development of a neural retriever at Malt that enables clients to instantly connect with their ideal freelancers.

First, we'll unravel the principles of a neural retriever, explaining why for our use case we picked it over a lexical retriever (keyword based). Then discover the creation of our custom neural retriever, including its architecture that uses a pre-trained small language model (SLM) and a custom transformer to encode the structure of profiles and projects. We’ll cover the training data and loss function that enable our latent space to representent the skill fit in a multilingual setting. We'll also examine the offline evaluation techniques we used to ensure its effectiveness. [12 mins]

Second, delve into the infrastructure that supports our neural retriever, including the reasons behind our choice of Qdrant as our VectorDB. Discover how we optimized the retriever encoder models by leveraging quantization and graph optimization techniques from ONNX to reduce latency and resource consumption (CPU deployment). [10 mins]

Finally, witness the transformative impact of our deployed retriever in production. We'll cover its impact on recommendation quality, user experience, and overall system performance. [3 mins]

No prior knowledge on recommender systems required. The first part requires some basic knowledge about deep learning but will not contain too much maths or code.

Marc Palyart is the Director of Data Science at Malt, the freelancer marketplace, where he leads the search and matching team. With over a decade of data-wizardry under his belt, he's ventured into the depths of academia and scaled the heights of industry where he's had the pleasure of collaborating with some truly remarkable people.

With a PhD in computer science, I completed my thesis in machine learning applied to the fashion industry in partnership with Lectra. I'm currently in charge of advanced matching topics requiring R&D at Malt, the freelance marketplace.