Keynote 28 - The Last Mile of Model Accuracy: Agent-Driven Federated Learning for Unseen Domains
Every AI/ML vendor ships a model that hits its benchmark and then watches it underperform in the field. The customer’s domain is different from the training distribution, and it keeps moving. It shows up everywhere — medical imaging, fintech, anti-fraud, identity and many others — and today it’s closed by guesswork: hunt for more data, retrain blind, repeat over weeks. The vendor is diagnosing a domain it cannot see. I’ll show what changes when an agent runs where the data already lives: it diagnoses what actually shifted and turns that into a targeted federated fine-tuning plan, without the vendor losing its own baseline. Two real deployments ground it — one in medical technology, one in anti-spoofing for identity — plus what it takes to make this work, and where it still breaks.
