Building Feedback-Driven Annotation Pipelines for End-to-End ML Workflows – February 18, 2026

​What you’ll learn:

​In this session, we’re going to build a complete curate-annotate-train-evaluate loop. We’ll focus on the specific logic that prevents you from wasting budget on the wrong data.

​You’ll leave with the code to:

• ​Label fewer images for similar performance: Random sampling is inefficient. We’ll use zero-shot selection and embeddings to mathematically identify the most unique samples in your dataset and allow you to maximize model coverage with a fraction of the usual labeling budget.

• ​Speed Up QA: You’ll learn to annotate and validate labels directly within FiftyOne, and use patch views to review specific objects and fix errors orders of magnitude faster than standard review.

• ​Build a hybrid data selection strategy: Most pipelines either label randomly (which is inefficient) or only chase failure cases (which makes the model forget normal cases). We’ll implement a balanced 30/70 split: 30% for diversity and 70% for targeting specific errors.

• ​Fix your data splits: It’s easy to accidentally cheat on your metrics. We’ll set up a rigorous workflow with a frozen test set for final scores and a separate golden set to catch label drift.

• ​Debug with embeddings: Model performance metrics don’t necessarily tell you what to fix. We’ll use embeddings to visualize the specific clusters confusing your model so you can target those exact scenarios.

​The Result: You’ll have a repeatable pipeline that helps you improve model performance with fewer labels, rather than just throwing more data at the problem.

​Who should attend:

​ML engineers, data scientists, AV/ADAS and robotics practitioners, computer vision researchers, data platform and MLOps engineers, and technical leads responsible for labeling, developing multimodal datasets and models, or maintaining consistent label semantics across projects and tools.