Introduction

This blog walks through the end‑to‑end architecture of Kermit, explaining how a modern agentic AI platform, cloud‑native services, and SAP Commerce integrations can come together to deliver secure, scalable, and intelligent customer interactions.

What is Kermit?

Kermit is a conversational AI assistant embedded within the client’s B2B Mobile Application. It enables customers to:

• Ask questions about products and pricing
• Place and manage orders using natural language
• Receive personalized recommendations
• Access account, property, and order‑related information
• Kermit has been designed with a code‑first, multi‑agent architecture and is fully deployed on AWS, ensuring reliability, scale, and strong governance from day one.

Business Use Cases Covered by Kermit

The Kermit Conversational AI platform supports a broad set of business and operational use cases across the digital ecosystem.  Below are few of the key use cases:

Customer Self‑Service & Support

• Order status inquiries (by order number, phone number, or account context)
• Order cancellation and modification requests
• Repair status checks and repair initiation
• Account management (profile details, properties, preferences)
• FAQs and policy‑driven questions (returns, terms & conditions, promotions)

Product Discovery & Commerce Enablement

• Guided product search using natural‑language queries
• Product availability, pricing, and packaging inquiries
• Personalized product recommendations
• Promotion and offer discovery
• Cart creation and pre‑checkout assistance

Analytics, Insights & Optimization

• Identification of order drop‑offs and escalations
• Measurement of bot‑assisted versus human‑assisted resolutions

Knowledge & GenAI‑Driven Assistance

• Knowledge Q&A using curated enterprise content (FAQs, product catalogs, policies)
• Context‑aware response generation using LLMs with enterprise guardrails
• AI‑generated conversation summaries for review and analysis

Architectural Overview

Below represents inferred reference architecture aligning Kermit components with enterprise CX and AI patterns. The inferred architecture assumes Claude Haiku 4.5 on Amazon Bedrock as the underlying LLM for Kermit.

1. Mobile Front End (iOS & Android Applications)
2. Security & Identity Layer (Microsoft ADFS, JWT)
3. Network & Load Balancing Layer (NLB + ALB)
4. Agentic AI Layer (Orchestrator + Sub‑Agents)
5. Data & Caching Layer (Redis, DynamoDB)
6. SAP Commerce & Enterprise Integrations
7. Monitoring, QA, and Analytics Layer

Each layer is independently scalable and designed following cloud‑native and zero‑trust principles.

Front End:  Mobile Application

The Mobile Application is the primary entry point for Kermit and is built as a fully native mobile experience:

• iOS: Swift & SwiftUI
• Android: Kotlin & Jetpack Compose

The chat interface communicates with backend services using Server‑Sent Events (SSE), enabling real‑time streaming responses from the AI in small chunks. This creates a responsive, chat‑like experience even when responses involve multiple backend calls or reasoning steps.

Security & Identity Management

Security is enforced and can be aligned with the respective client’s enterprise standards:

• Microsoft ADFS acts as the centralized Identity Provider
• Users authenticate once via the mobile application
• JWT tokens are issued post‑authentication
• Each chat request carries:
  • An Authorization Token (standard ADFS access token)
  • An X‑Client Token carrying customer and contextual data

Kermit itself never issues authentication tokens. Instead, it validates tokens with the client’s mobile platform, maintaining strict separation of concerns.

Network & Load Balancing

All inbound traffic flows securely through:

1. Network Load Balancer (NLB) – external entry point hosted in the client’s AWS account
2. Application Load Balancer (ALB) – internal routing within a private VPC

From here, requests are routed to containerized services running on Amazon ECS with AWS Fargate. All components operate inside a private VPC, with AWS VPC Endpoints providing secure, private access to services such as Bedrock, S3, and CloudWatch.

Prefetch Service: Reducing Latency from First Interaction

To improve perceived performance, Kermit introduces a Prefetch Service:

• Implemented as an AWS Lambda function
• Triggered immediately after user login
• Pre‑loads:
  • User‑specific SAP Commerce data
  • Frequently accessed product information

This proactive cache warming significantly reduces cold‑start latency during the user’s first conversational interaction.

The Agentic AI Layer: Where Intelligence Lives

Orchestrator

At the core of Kermit is a stateless Orchestrator built using:

• FastAPI (Python 3.13)
• Uvicorn ASGI server

Responsibilities include:

• Managing chat sessions
• Interpreting user intent using Claude Haiku 4.5 via Amazon Bedrock
• Routing requests for specialized sub‑agents
• Streaming responses back to the client via SSE
• Enforcing rate limits and health checks

All conversation state persisted externally in DynamoDB, keeping the orchestrator horizontally scalable.

Sub‑Agents (Agent Core)

Kermit uses multiple domain‑specific AI agents, each deployed as a separate ECS Fargate task and running on Amazon Bedrock Agent Core using the Strands Agent Framework:

• 1. Account Info Agent – customer profiles, properties, contacts
  2. Order Agent – full order placement workflows
  3. Order & Recommendations Agent – order history and recommendations
  4. Product Info Agent – catalogue, pricing, fuzzy search

Agents communicate securely using SigV4‑signed HTTP requests and coordinate via an Agent‑to‑Agent (A2A) protocol.

Bedrock, Guardrails, and RAG

The conversational reasoning across Kermit is powered by Claude Haiku 4.5 via Amazon Bedrock.

Key capabilities include:

• Bedrock Guardrails to enforce content safety and fertilizer‑domain relevance
• Retrieval‑Augmented Generation (RAG) using Amazon Bedrock Knowledge Bases
• Source content stored in Amazon S3 (PDFs, CSVs, structured documents)

This ensures responses are both accurate and grounded in client‑approved data.

Data Layer & Caching Strategy

Caching:

• Amazon ElastiCache Serverless (Redis) is used for caching.
• A dedicated Cache Manager Lambda controls read/write/invalidation logic

Cache tiers include:

• User‑specific SAP Commerce data (TTL: 1 hour)
• Global product & pricing data (TTL: 24 hours, refreshed every 4 hours)

Transactional operations (order placement, cart submission) always bypass the cache to guarantee data consistency.

Conversation History

All conversational states are stored in Amazon DynamoDB:

• Orchestrator chat history
• Sub‑agent Conversation
• Session validity enforced within an 8‑hour window

SAP Commerce & Enterprise Integrations

Kermit integrates deeply with enterprise systems via:

• SAP Commerce APIs exposed through SAP API Management
• B2B Mobile Application APIs for business operations
• ArcGIS mapping integration for geo‑spatial and property‑based context

This allows the AI to operate with full business awareness rather than acting as a standalone chatbot.

Observability, QA, and Analytics

Monitoring

• Amazon CloudWatch for logs and metrics
• Centralized error processing using Lambda, DynamoDB, Event Bridge, and SES
• Hourly error digests sent to engineering teams

Quality & Evaluation

• Automated LLM evaluations using DeepEval
• Deployment‑triggered regression testing
• Dashboards via Bedrock Analytics and Amazon QuickSight

Deployment Model

All components can be deployed via CloudFormation across:

• Dev
• QA
• Production

Compute choices are optimized per workload:

• ECS Fargate – APIs and AI agents
• AWS Lambda – prefetch, caching, error handling

Closing Thoughts

Kermit represents a shift from traditional rule‑based digital assistants to a true agentic AI platform—secure, scalable, and deeply integrated with enterprise commerce systems.

By combining multi‑agent orchestration, modern LLMs, and cloud‑native architecture, one of our clients has laid the foundation for intelligent, context‑aware customer engagement in the agriculture sector—with clear pathways for future expansion into voice and advanced multimodal experiences.