How Capital One Delivers Multi-Agent Systems [Rashmi Shetty] - 765

TWiML AI (This Week in Machine Learning & AI)
TWiML AI (This Week in Machine Learning & AI)Apr 16, 2026

Why It Matters

Capital One’s agentic AI platform shows how regulated enterprises can safely scale generative AI, turning complex customer interactions into automated, compliant workflows that drive speed and personalization.

Key Takeaways

  • Capital One shifted from classic ML to LLM-driven action systems.
  • Multi‑agent architecture breaks complex goals into specialized, orchestrated agents.
  • Chat Concierge pilots agentic AI for auto‑dealer customer matchmaking.
  • Built‑in governance, risk, and policy layers ensure regulatory compliance.
  • Platform provides SDKs and tools for rapid, safe agent development.

Summary

The TWIML AI podcast episode features Rashmi Shetty, senior director of Capital One’s enterprise generative AI platform, explaining the bank’s transition from traditional machine‑learning pipelines to large‑language‑model (LLM) driven systems that can actually execute actions. She outlines how the organization moved from simple response generation to a multi‑agent framework that decomposes large, goal‑oriented problems into discrete tasks handled by specialized agents. Key insights include the rationale for multi‑agent architectures: complex use cases require step‑wise orchestration, with each agent responsible for a narrow function such as intent disambiguation, planning, risk validation, or user‑facing response formatting. The flagship "Chat Concierge" pilot demonstrates this approach in the auto‑dealer space, matching customers to vehicles and automating follow‑up actions like test‑drive scheduling. Shetty emphasizes that regulatory compliance is baked into both the platform and individual agents via policy guards, risk‑office oversight, and evaluation gates. "Policy‑bound agentic operations" ensure that agents cannot violate banking regulations or expose the bank to cyber risk, while still delivering personalized experiences. The broader implication is a new developer experience: Capital One’s internal platform supplies SDKs, memory services, data‑lineage tools, and latency controls, enabling rapid, secure deployment of agentic solutions at scale. This positions the bank to leverage its data advantage while meeting stringent financial‑industry standards.

Original Description

In this episode, Rashmi Shetty, senior director of enterprise generative AI platform at Capital One, joins us to explore how the company is designing, deploying, and scaling multi-agent systems in a highly regulated environment. Rashmi walks us through Chat Concierge, a multi-agent chat experience for auto dealerships that handles intent disambiguation, tool invocation, and human handoffs to deliver safer, more personalized customer journeys. We discuss Capital One’s platform-centric approach to AI agents and how it separates design from runtime governance, embedding policies, guardrails, and cyber controls across agent threat boundaries. Rashmi shares how the team approaches the developer experience for agent builders, observability, and evals for stochastic, multi-agent workflows; and strategies for model specialization, including fine-tuning and distillation. We also cover standards and abstraction, closed-loop learning from production telemetry, and key lessons for enterprises building agentic systems.
🗒️ For the full list of resources for this episode, visit the show notes page: https://twimlai.com/go/765.
🔔 Subscribe to our channel for more great content just like this: https://youtube.com/twimlai?sub_confirmation=1
🗣️ CONNECT WITH US!
===============================
Subscribe to the TWIML AI Podcast: https://twimlai.com/podcast/twimlai/
Follow us on Twitter: https://twitter.com/twimlai
Join our Slack Community: https://twimlai.com/community/
Subscribe to our newsletter: https://twimlai.com/newsletter/
Want to get in touch? Send us a message: https://twimlai.com/contact/
📖 CHAPTERS
===============================
00:00 - Introduction
04:05 - Motivation for multi-agentic systems
05:33 - Chat Concierge
09:26 - Regulations
11:19 - Governance: agent building vs. agent runtime execution
12:26 - Purpose of the enterprise platform
13:50 - Policy-bound agentic operations
15:30 - Developer requirements for agentic systems vs classic ML/AI architecture
19:40 - Observability for agentic systems
22:36 - Observability challenges in multi-agentic workflows
25:33 - Agentic evaluation frameworks vs. classic ML evaluation frameworks
27:50 - Reasoning and specialized models
29:40 - Key architectural decisions
32:23 - Agentic standards
33:33 - Human-in-the-loop integration
38:23 - Managing complexity for teams
43:36 - Model risk office
45:02 - Lessons learned
48:13 - Next steps
🔗 LINKS & RESOURCES
===============================
Evolving MLOps Platforms for Generative AI and Agents with Abhijit Bose - #714 - https://twimlai.com/podcast/twimlai/evolving-mlops-platforms-for-generative-ai-and-agents
🎙️Microphone: https://amzn.to/3t5zXeV
🎛️ Audio Interface: https://amzn.to/3TVFAIq
🎚️ Stream Deck: https://amzn.to/3zzm7F5

Comments

Want to join the conversation?

Loading comments...