Embedding models are a foundational component of modern ML systems, enabling clustering, ranking, and large-scale information retrieval. Despite their widespread use, their training dynamics and evaluation are often treated as a black box, which makes adapting them to real production needs difficult.
This session explains how embedding models are trained, focusing on representation learning theory, common contrastive and metric-learning loss functions, and the data preparation strategies that shape embedding quality. A short code walkthrough inspired by an EmbeddingGemma-style training pipeline illustrates dataset construction, pairing, and negative sampling in practice. We will connect these training choices to downstream retrieval performance and discuss how to evaluate embeddings for a concrete application, using lessons learned from improving a agentic search application at Dell.
In this session you will learn how embedding objectives affect real-world performance, how to diagnose failure modes, when fine-tuning is justified. The goal is to provide a practical mental model of embedding systems that engineers can apply directly in production.
Main Takeaways:
- How embedding training actually works: contrastive objectives, negatives, and why most pipelines fail silently
- How to evaluate embeddings for real systems: beyond cosine similarity into retrieval quality and business impact
- When to fine-tune and when not to: practical signals, tradeoffs, and production lessons learned
Interview:
What is the focus of your work these days?
I am part of an AI solution teams that develop internal solutions at Dell. In the past 2 years I lead the research and development of the agentic search app for Dell employees, with a focus on the multi agentic architecture, evaluation and model training to improve quality. My work spans the full lifecycle from data curation and training to evaluation and deployment, with an emphasis on systems that deliver measurable impact in real engineering workflows.
What is the motivation behind your talk?
Embedding models are everywhere, but most teams treat them as plug and play components without understanding how training choices affect performance. I wanted to break open that black box and share practical lessons from production systems, where small changes in data and objectives make a significant difference. and generally speaking I always like to share my learnings.
Who is your talk for?
Engineers, data scientists, and technical leaders building or operating ML systems that rely on search, retrieval, or ranking. No deep prior experience with embedding training is required, but familiarity with ML concepts will help.
Speaker
Rachel Shalom
AI Applied Scientist & Distinguished Engineer @Dell, 7+ Years Experience Turning Cutting Edge AI into Real World Impact
I’m an applied AI scientist and Distinguished Engineer with 7+ years of experience turning cutting edge AI into real world impact. I specialize in LLM-based knowledge retrieval, multi-agent systems, and deep learning for time series data, building solutions that don’t just work, they get used.
I lead end to end, from data curation and plumbing to prompt engineering, model training, evaluation, and production deployment. I operate at the intersection of hands-on engineering and strategic leadership, partnering across engineering and product to deliver systems that move the needle.
I hold a B.Sc. in Mathematics, an MBA, and graduated with honors from Yandex’s elite machine learning program. Always curious, always shipping.
