Jouanneau & Palyart - SLM-powered retrieval to scale freelancers matching at Malt

MALT is Europe’s leading freelancing marketplace with over 700,000 freelancers, focusing on matching freelancers with companies through an AI-powered system

The team developed a hybrid retrieval system combining:

• BM25 for lexical matching
• Neural retrieval using small language models
• Approximate Nearest Neighbors (ANN) for efficient large-scale matching

Key architectural components:

• Multilingual BERT model as frozen backbone (113M parameters)
• Trainable head (7M parameters)
• Section-aware encoding to preserve document structure
• Skip connections and weighted sum pooling

Performance improvements:

• Reduced execution time from tens of seconds to max 3 seconds
• Maintained precision while improving latency
• Successfully implemented cross-lingual matching
• Reduced model size through quantization

Model training approach:

• Used contrastive learning with triplet loss
• Leveraged job taxonomy for latent space organization
• Maintained language alignments through careful architecture choices
• Augmented training data with historical interactions

Production considerations:

• Vector database implementation for efficient retrieval
• Regular reindexing to handle profile updates
• CPU-only deployment for cost efficiency
• Balanced trade-off between precision and retrieval speed

Built-in support for:

• Cross-lingual matching
• Geospatial filtering
• Job category organization
• Semantic understanding across languages