Building an AI Platform: Memory, Tools, Guardrails, Observability & Shipping a Production Agent
Orchestrating Safe and Stateful AI Applications - Dewang Sultania
A functioning AI platform requires much more than just routing queries to a model. Real world applications need memory to maintain context, tools to take action, and strict safety policies to prevent costly mistakes.
In the second episode of our upcoming series featuring the authors of Designing AI Systems, Dewang Sultania will break down the final steps of deploying a production grade AI agent. We will explore how to manage session memory, enforce guardrails as execution policies rather than basic filters, and build proper observability to track system health. We will also discuss the critical shift from building a simple chatbot to an agent where every executed action has permanent consequences.
He’ll cover:
Session service and memory management
Tools and MCP
Guardrails as execution policies rather than just filters
Observability for AI systems
Workflow orchestration and deployment
About the Speaker:
Dewang Sultania is a Senior Machine Learning Engineer at Netflix, where he designs scalable systems for multimodal generative AI, diffusion models, and video processing. Previously at Adobe, he built production systems for large language models, including data pipelines, fine-tuning, retrieval-based systems, and prompt engineering. He also helped design the platform infrastructure used to deploy large language models across Adobe's product suite.
DataTalks.Club is the place to talk about data. Join our Slack community!