Demo to Production. Architect a Real Agentic AI System (Step by Step)

Step-by-step guide to architecting a production-ready agentic AI system for banking customer support with security and domain-specific agents.

Ask about this video. Answers come from its transcript only — with the timestamp, so you can check them.

Generated from the transcript and can be wrong — check the timestamp.

Key Takeaways

Agentic AI systems can be architected step-by-step from simple language models to complex multi-agent systems.
Security and privacy are critical, requiring token-based authentication and PII reduction.
Domain-specific sub-agents improve response accuracy and system modularity.
Conversation history management is essential for contextual and secure customer interactions.
Robust infrastructure and network security are necessary for production readiness in banking AI applications.

What the video covers

Introduction to building a production-level AI application for banking customer support.
Initial use of language models for basic queries like account balance with privacy considerations.
Development of domain-specific sub-agents such as accounts agent, transaction agent, and customer service agent.
Integration of multiple agents to handle complex user requests involving balance and transaction details.
Explanation of MCP (Multi-Component Processor) architecture for managing agents in a scalable system.
Implementation of security measures including token-based authentication and identity providers.
Handling privileged customer requests with role-based permissions in AI interactions.
Management of conversation history and stateless interactions for accurate and secure responses.
PII (Personally Identifiable Information) reduction and secure handling of sensitive data.
Overview of production-ready infrastructure including network security, firewalls, and API gateways.

Chapters

What is the role of domain-specific sub-agents in this AI system?

Domain-specific sub-agents handle specialized tasks such as account balance inquiries, transaction details, and customer service, enabling modular and accurate responses.

How does the system ensure security and privacy for banking customers?

The system uses token-based authentication, integrates with existing identity providers, and implements PII reduction to protect sensitive customer information.

What is MCP architecture and why is it important?

MCP (Multi-Component Processor) architecture manages multiple agent servers for different domains, allowing scalable, modular, and production-ready AI system deployment.

Full Transcript — Download SRT & Markdown

Speaker A

Hey guys, production level AI application software engine. So, let's get started. Thanks for all your support and love.

Speaker A

So, I'm very happy about that, and we have a long way to go, banking organization customer support team.

Speaker A

Last month, customer support team 4.2 2K 65% of the customers contacting organization and the network provider.

Speaker A

Which means in the organization millions of dollars already customers banking more than 14 screens.

Speaker A

So bank build a secure customer support customer support. So step by step. So first question language model basic: What is my account balance? Response: Sorry, I don't have access to your bank accounts or account balances, which is valid.

Speaker A

So step two. So next internal transaction details, statement request, change of addressing. So tool-based agent: What is my balance? Question language response, for example, your account balance is so and so. Natural language response tools banking applications tool overload.

Speaker A

So step three: create domain-specific sub-agents, accounts agent, transaction agent, service agent, agent customer agent, customer service.

Speaker A

So for example, user question: What is my balance and can you also get me my latest transactions? Agents, right. Step four complete system.

Speaker A

So if user: Get me my balance along with the last five transactions. Right. Right.

Speaker A

So application but still in the—for example, service tightly coupled programs manage agent. So next different MCP server, accounts MCP server, transaction MCP server, service MCP server, and MCP server. So MCP architecture MCP explained, right? MCP design.

Speaker A

So, what is my account balance? [snorts] So, balance inquiry. So balance your customer, your savings account balance is 24,345. Please give me your customer.

Speaker A

So application application security authentication. Next step, right, evident existing identity provider. So browser first password.

Speaker A

So and most probably token-based authentication authentication. So first level of securityen welcome. Can you increase my credit limit to five lacs first customer? Yes, I can help you with that.

Speaker A

Thanks for confirming your credit limit has been increased. Second customer, sorry, I do not have permissions to increase your credit. So in the functionality next privileged customers, premium customers, standard customers, privileged customers, credit card service agent increase.

Speaker A

AI software software engineering based software authentic application. Now first, what was my last transaction?

Speaker A

Your last transaction was a debit of 200 to Amazon suspicious. Was that the one I flagged as suspicious last week?

Speaker A

I don't have any record of a transaction being flagged. Could you clarify which transaction you mean? So in the response first statement and second statement could you clarify which previous, which means complete conversation history are stateless.

Speaker A

Okay. So in the past conversation history conversation history conversation history agent shared state. Example: debit limit to five lakhs. My credit card number is credit card number.

Speaker A

So customer personally identifiable information. So first complete credit card number service right so the different components. First one PI reduction. So user PI reduction service personally identifiable information again.

Speaker A

For open self-hosted so that third party only for complex reasoning system and self-hosted so that information system before submitting a checkbook request always confirm the delivery address with the customer. So submit the checkbook request using customers registered details.

Speaker A

Agent evaluation. So in the evalu nondeterministic flow right natural language flow different what's Account balance agent confident your account balance is 18,200.

Speaker A

So customer bank 52,340 but agent 18,200 so customer support 342 342 customers us. So observability green and blue general software system CPU memory discular interaction Fore system.

Speaker A

So 100% perfect software maintenance. So rightet applic control. So network infrastructure network. So in the US banking layer security control web application firewall denial of service rate limits limits API gateway AP security edge layer security infrastructure aspect.

Speaker A

So in the architecture already completed production ready architecture. So I'll see you in the next video.

Speaker A

Thanks for watching. Bye-bye.

Topics:Agentic AIBanking AICustomer Support AIMulti-Agent SystemMCP ArchitectureToken-Based AuthenticationPII ReductionAI SecurityConversational AIProduction Ready AI

Get More with the SozAI App

Transcribe recordings, audio files, and YouTube videos — with AI summaries, speaker detection, and unlimited transcriptions.

App Store Google Play

Or transcribe another YouTube video here →

Free tools: TXT to SRT · SRT Validator · Merge SRT · Subtitle to Text · All tools