Rethinking Fine-Tuning: Building Adaptive Pipelines on Top of Pretrained Models

**Abstract** What if your model isn’t the end of the pipeline—but the beginning of a smarter one? In this talk, we’ll explore building a lightweight, modular adjustment layer on top of pre-trained models, designed to inject strategic signals into your system: business objectives, domain expertise, blind spot corrections, or patterns the model wasn’t even trained on. The result is a flexible, ensemble-style architecture that adapts without retraining the core model.This approach gives you an easy way to fine-tune model behavior in production to reflect contexts, constraints, or objectives that weren’t explicitly modeled during training. **About the Speaker** Racheli is a Data Scientist at Riskified with 5 years experience in the Fraud Detection domain. She has led multiple projects regarding automating and optimizing ML solutions, and is currently working on data sampling related projects. Outside of her professional life, Racheli enjoys traveling, camping and Yoga.