Stadtwerke Concierge

The Challenge

Munich's municipal utilities company (Stadtwerke) handles thousands of customer inquiries daily. Traditional customer service requires extensive manual work for repetitive tasks like moving notifications, meter readings, and emergency handling. This creates bottlenecks, increases costs, and leads to inconsistent customer experience.

The Solution

Stadtwerke Concierge is an AI-powered chatbot that automates the entire customer service workflow. Using AWS Bedrock's Claude API, it understands complex German-language requests, manages multi-step conversations, and provides accurate information for three core use cases: moving services, emergency situations, and meter readings.

Technical Architecture

System Flow

User Request → Spring Boot API → PostgreSQL Session Management → AWS Bedrock Claude → Response Generation

Frontend: Clean REST API interface for easy integration
Backend: Spring Boot with comprehensive error handling and logging
Database: PostgreSQL for conversation history and user sessions
AI Engine: AWS Bedrock Claude for German language understanding
Deployment: Docker containers on AWS App Runner

Tech Stack

Language: Java 17+
Framework: Spring Boot 3.x with Spring Security
Database: PostgreSQL with Hibernate ORM
Testing: JUnit 5, Mockito, Spring Test (85%+ code coverage)
AI Services: AWS Bedrock (Claude API)
Containerization: Docker & Docker Compose
Deployment: AWS App Runner, CloudWatch Logs
CI/CD: GitHub Actions (automated testing)

Key Features

🔄 Multi-Turn Conversations

Maintains complex conversation context across multiple exchanges. The system understands references to previous messages and can clarify ambiguous requests through intelligent follow-ups.

🇩🇪 German NLP Support

Native German language understanding via AWS Bedrock Claude. Handles colloquialisms, regional variations, and complex sentence structures specific to German customer interactions.

📋 3-Phase Service Flow

Moving Services: Guides users through address changes, service connections. Emergencies: Rapid response for gas leaks, outages. Meter Readings: Automated data collection and validation.

💾 Session Persistence

All conversations stored in PostgreSQL. Users can resume sessions, view history, and receive consistent assistance across multiple interactions.

🛡️ Error Handling & Validation

Graceful error recovery, user-friendly fallbacks, and comprehensive logging. System degrades gracefully if AWS Bedrock is unavailable.

📊 Analytics & Monitoring

CloudWatch metrics for conversation success rates, response times, and error tracking. Real-time monitoring dashboards for operations team.

Results & Metrics

94.4%

Success Rate

Across 18 comprehensive test cases

85%+

Test Coverage

Unit + Integration tests with TDD

<500ms

Avg Response Time

API to Claude to User

Core Service Flows

Moving, Emergency, Meter Readings

Key Learnings & Technical Insights

🧪 Test-Driven Development at Scale

Implementing TDD from the ground up resulted in 85%+ code coverage and significantly reduced bugs. Testing AWS integration points (mocking Bedrock) required creative patterns.

🤖 AI Integration Challenges

Working with AWS Bedrock Claude revealed the importance of prompt engineering and conversation context management. Initial responses were inconsistent until we implemented context windowing.

🇩🇪 Language-Specific Development

German NLP is significantly different from English. Handling edge cases like "Umlaut" support, regional dialects, and complex grammar required extensive testing with native speakers.

☁️ AWS Bedrock Integration

First-hand experience integrating cutting-edge Claude API. Learned proper token management, cost optimization, and handling API rate limits for production workloads.