← All speakers

Bio, Work & Ideas

Miguel Martinez

Conference affiliation: Microsoft · 2024

Miguel Martinez is a principal AI solution engineer at Microsoft who helps businesses build AI applications grounded in their own data, security requirements, and operational needs. His work encompasses healthcare data science, enterprise copilots, AI agents, and multimodal systems.

Originally from Venezuela, Martinez moved to the United States with his family during high school. He spent seven years at UnitedHealth Group, becoming a principal data scientist at Optum focused on healthcare challenges including pharmacy operations and hospital admissions. He subsequently joined Microsoft as a senior technical specialist for data and AI before becoming a principal AI solution engineer.

His technical priorities include:

  • Customer-scoped enterprise AI: At AI Engineer World’s Fair 2024, he supported a retail copilot workshop combining Azure AI Search with customer data in Azure Cosmos DB. He emphasized establishing authenticated customer identity before generating personalized answers, limiting opportunities to expose another customer’s information through conversational manipulation.
  • Spanish-language AI agent education: He led a Microsoft Reactor session on Semantic Kernel covering plugins, memory, external systems, and agent workflows.
  • Multimodal business applications: At Microsoft AI Tour Houston, he led a workshop on multimodal models in Azure AI Foundry, expanding his practical instruction beyond text-based assistants.

Read the topics behind these talks

1 conference talk

Key ideas

Scroll to read ↓

Follow a retail RAG backend from product retrieval and customer lookup through prompt assembly, deployment and an evaluation that catches a convincingly invented product.

  • What information does a retail chatbot need?
    0:00 ↗
  • A tent recommendation, then a purchase-history question
    7:07 ↗
  • Two retrieval paths meet in one prompt
    10:55 ↗
  • Launch the lab, authenticate and inspect the data
    16:22 ↗
  • What a flow owns
    40:13 ↗
  • Prompt flow, application orchestration and the playground
    49:48 ↗
  • Connections, persistence and search quality
    1:04:31 ↗
  • Customer context is not customer authorization
    1:17:25 ↗
  • From lookup results to a rendered prompt
    1:26:51 ↗
  • Run locally, package the flow and understand its limits
    1:33:50 ↗
  • A fluent answer can still invent a product
    1:41:20 ↗
  • Use failures to change the system
    1:49:33 ↗
  • Uploaded PDFs need a retrieval lifetime
    1:52:02 ↗

References