Foundry demosFine-tuning / Evaluation / Agent improvement
Turn agent traces into a reinforcement fine-tuning loop
Use a retail support scenario to explore traces, a business-specific reward function, and a specialized model evaluated against a baseline.
In this demo
A retail refund-and-exchange agent demonstrates how operational traces can feed an improvement loop. Jeff inspects tool trajectories, explains a reward function covering decisions, financial amounts, and workflow quality, then compares a specialized smaller model with a baseline on example support requests. The emphasis is on capturing business-specific feedback and evaluating both quality and economics.