Turn your data into AI advantage with Training in Microsoft Foundry | BRK252
Simon Zhao, Shengtao Ma, Vijay Aski, and Meng Tang present Microsoft Foundry’s training capabilities and examine how production data and agent task feedback can guide model improvements. The session combines a platform overview, a code review optimization demo, and customer examples.
Training models for business-specific tasks
The session focuses on adapting broadly available models to business requirements. It examines when to use training and introduces three approaches:
- Fine-tuning to customize model behavior for specific tasks.
- Direct preference optimization (DPO) as a post-training technique.
- Reinforcement learning (RL) to improve performance through task feedback.
The presenters discuss how production data and agent experience can feed continuous model improvement, with attention to accuracy, agent performance, token efficiency, and deployment costs. Customer examples illustrate accuracy gains and cost reductions, although the session description does not provide numerical results.
Microsoft Foundry training capabilities
The platform overview covers managed training, programmed training, and reinforcement learning environments. It also introduces expanded fine-tuning support for open models and Microsoft MAI models.
A deeper technical segment examines post-training and RL environments, followed by a code review optimization demonstration. The session closes by discussing how trained models are deployed to production through Foundry.
Customer case studies
Docusign: Agreement information extraction
Docusign’s case study examines fine-tuning and reinforcement learning for agreement AI extraction. It shows how customized open models with Fireworks on Foundry can support task-specific accuracy and production use.
Heidi Health: Medical AI
Heidi Health’s case study covers medical AI training, reinforcement learning, and cost-efficient deployment. It also describes the use of customized open models with Fireworks on Foundry.
Session chapters
- 00:00 — Introduction to turning data into a competitive advantage with AI training
- 00:22 — Why training matters for AI agents and an overview of Foundry
- 01:14 — Customer examples of accuracy gains and cost reductions through model training
- 04:18 — When to use training; fine-tuning, DPO, and reinforcement learning
- 05:51 — Managed training, programmed training, and RL environments
- 09:20 — Expanded fine-tuning support for open models and Microsoft MAI models
- 10:16 — Docusign agreement AI extraction case study
- 19:48 — Post-training, RL environments, and code review optimization demo
- 40:09 — Heidi Health medical AI training and deployment case study
- 46:52 — Deploying trained models to production with Foundry
Resource
This is Microsoft Ignite 2026 session BRK252, an intermediate-level Agent Platform breakout presented in English (US).