Rubberduck – AI-Powered Developer Assistant

Executive Summary

Scania's development team needed a faster way to navigate their extensive GitLab repository ecosystem and internal documentation. Developers were spending significant time searching for code examples, understanding project structures, and finding internal best practices across multiple repositories.

Awiant delivered a custom AI-powered assistant built on Strands SDK and AWS Bedrock that provides instant, context-aware answers about code, repositories, and internal processes. The solution integrates seamlessly with GitLab and platform Wiki, enabling developers to ask natural language questions and receive accurate, source-cited responses. The assistant runs on Anthropic's Claude models via AWS Bedrock.

About the Customer

Scania PDM is the Product Data Management organisation within Scania, responsible for developing and operating IT solutions that support product development and lifecycle management. The organisation builds and maintains internal tools, platforms, and integrations used by engineers and developers across Scania and Traton.

Scania is a leading manufacturer of trucks, buses, and engines, with a strong focus on digital transformation and software development.

Challenge

Scania's PDM development teams faced several productivity obstacles:

• Time-consuming onboarding – New team members struggled to understand the codebase and internal processes
• No unified search – Developers had to manually browse multiple GitLab repositories to find code examples or understand project structures
• Knowledge fragmentation – Project-specific conventions and best practices were scattered across repositories and wiki pages
• Context switching – Developers frequently switched between GitLab, wiki, and documentation to answer questions
• Fragmented business rules – a single rule could be spread across a canonical index and several interlinked spreadsheet sources, requiring manual cross-referencing to understand what a relation actually means

Solution

Awiant designed and implemented PDM IT Rubberduck, a custom AI assistant built on AWS Bedrock using Anthropic's Claude model with the Strands SDK. The solution uses a serverless architecture where users interact with a React frontend that connects via WebSocket to Lambda functions running the AI agent.

The solution itself was largely co-developed with Claude Code, Anthropic's agentic coding tool — the same AI technology that powers the assistant was used to build it.

Technologies Used

AWS Bedrock + Claude, Strands SDK, AWS AgentCore Memory, Lambda + DynamoDB, React + TypeScript, Microsoft Entra ID, Terraform, GitLab API, Claude Code (development).

How It Works

The assistant is built with the Strands SDK and leverages MCP-style tool use via AWS Bedrock tool calling to reason iteratively: it calls tools, inspects results, updates its understanding, and decides the next best action. This continuous decision loop allows the agent to navigate between high-level architecture questions and low-level code details across repositories and internal documentation.

1. Searches GitLab repositories – Finds relevant code and reads file content
2. Loads project context – Automatically detects and loads project-specific documentation
3. Searches dual sources – Combines information from both GitLab and platform Wiki
4. Consolidates fragmented sources – merges business rules spread across canonical and compatibility sources, cross-references relations against object-type definitions, and keeps documented facts, source statements, and inference clearly separated
5. Maintains conversation memory – Uses AgentCore Memory for personalized responses
6. Streams responses – Provides real-time feedback as answers are generated

Result

With Rubberduck in place, PDM IT developers can now find code and documentation much faster than before. The AI assistant automatically searches GitLab repositories and the platform Wiki, reads actual file content before answering, and cites sources with file paths and project names — eliminating hallucinations and ensuring accurate, context-aware responses.

• Over 100 active users – in full production at Scania PDM IT, with rollout to a wider audience under way
• Faster code search – Questions answered quickly with accurate, source-cited responses
• Reduced search time – Developers spend less time manually browsing repositories
• Faster onboarding – New team members productive faster with AI-guided code exploration
• Scalable architecture – Designed to handle multiple repositories and concurrent users
• Future-proof platform – Ready for S3 Vectors integration and enhanced AI capabilities