Presented by Vijay Aski, Chris Lauren, Ankit Parag Shah, and Wei Wei at Microsoft Ignite, this session explores practical approaches for training and deploying reasoning models using Azure ML and Microsoft Foundry.

Training and Deploying Reasoning Models with Microsoft Foundry and Azure ML

Overview

This session guides viewers through the process of training, hosting, and inferring custom reasoning models using Microsoft Foundry together with Azure Machine Learning (Azure ML). It’s designed for an intermediate audience looking to expand their expertise in applying large language models (LLMs) and reinforcement learning for business-focused AI solutions.

Key Session Topics

  • Real-World AI Adoption: Customers are leveraging AI Foundry for tangible ROI, deploying agent-driven applications in multiple industries.
  • From Prototype to Production: Addresses typical challenges faced when scaling AI models from experimental phases to real business environments.
  • Harnessing LLMs: Reference to research on using LLMs for actionable business insights, including demonstration of a financial analyst agent.
  • Reinforcement Learning Techniques:
    • Introduces reinforcement learning for improving model accuracy.
    • Explains both synchronous and asynchronous reinforcement learning methods.
    • Shows how reinforcement learning enables multitask and generalization capabilities in reasoning models.
  • Model Evaluation and Benchmarking:
    • Demonstrates the evaluation pipeline comparing multiple fine-tuning approaches.
    • Presents benchmarking results on throughput and token latency improvement.
  • Deployment Practices: Step-by-step demonstration of hosting models for real-time inference and integration into applications via Foundry and Azure ML.

Technical Focus Areas

  • Azure ML Integration: Training and deploying models using Azure’s managed ML service.
  • Microsoft Foundry: End-to-end reasoning model workflows, including agent creation and operationalization.
  • Model Fine-Tuning: Comparing generalization and multitask performance of several RL techniques.
  • Inference and Performance: Optimization for production workloads, measuring key metrics (throughput, latency).

Practical Takeaways

  • Guidance for building, fine-tuning, and deploying custom reasoning and agent models using Microsoft’s AI stack.
  • Insights on moving from prototype experiments to robust, scalable AI solutions compliant with real-world requirements.
  • Strategies for evaluation and improvement using reinforcement learning and benchmarking.
  • Direct application examples relevant for financial analysis agents and other business scenarios.

References & Resources

Speakers

  • Vijay Aski
  • Chris Lauren
  • Ankit Parag Shah
  • Wei Wei

Intermediate technical walkthrough—focused on actionable methods and learnings from deploying advanced AI reasoning models with Microsoft’s tools.