Building State-of-the-Art, Open-Source Embedding Models

Talk "Building State-of-the-Art, Open-Source Embedding Models" by Max Cembalest

Get this course for free with Premium Ai+ subscription

Description

Deep learning-powered retrieval systems are the backbone of modern LLM applications, enabling efficient semantic search, recommendation, and knowledge retrieval. This talk unpacks the journey behind building, training, and scaling the Nomic Embed model series—one of the most widely used open-source, multilingual, multimodal embedding families on Hugging Face, with over 35M+ downloads. We’ll explore the data curation pipeline, training infrastructure, and inference-time objectives that make these models state-of-the-art, including innovations like Matryoshka resizability, quantization, and sparse mixture-of-experts training. Additionally, we'll discuss the unique challenges of pushing applied deep learning forward in an industry lab, where real-world constraints drive cutting-edge advancements in embedding model design.

Instructor's Bio

Max Cembalest

Developer Advocate at Nomic AI

Max Cembalest is a developer advocate at Nomic AI who blends a rigorous math and CS foundation (Wesleyan B.A., 3.95) and an MS in Data Science from Harvard with five years building ML products and educational technology. He previously worked as an ML engineer at Arthur and taught mathematics and computer science at African Leadership Academy, using classroom practice to shape practical, teacher-centered AI tools. Max's hands-on experience ranges from Cozmo robotics and Scratch integration to shipping ML systems, giving him a rare mix of developer, educator, and product intuition. A self-described "creative quant" with a love of puzzles and a theatrical background, he brings clear communication, playful curiosity, and a healthy skepticism of undue faith in AI — plus the conviction that logarithms are underrated.