Martin KrausLeads DevOps for High-Performance Computers; co-authored the postAUMOVIO
Automotive embedded software. A hackathon, then a pilot build with AWS, using Cline in VS Code.
AWS for Industries blog (vendor case study), published Read the case study (opens AWS for Industries blog (vendor case study) in a new tab)
Observation
AUMOVIO's Cline trial fixed a five-day bug in minutes, but general models ignored the company's own libraries.
- Takeaway
- Agentic search finds code that one-shot retrieval misses, but a general model does not reuse your internal libraries unless something tells it to. Check pilot output for duplicates of existing APIs, and decide whether rules, MCP tools, or a tuned model carry that knowledge.
- What happened
- AUMOVIO ran a hackathon with AWS to test a coding assistant on its software-defined-vehicle code. Retrieval from a vector store could not find the right code in one query. So the team switched to Cline on Claude 3.7 Sonnet in Amazon Bedrock, which searches the codebase step by step. As anecdotes, the team reports Cline fixing in minutes a bug that had taken a senior developer five full days, and halving a very large file by removing redundancy. The same models, not tuned on AUMOVIO's code, wrote duplicates of existing internal libraries. The pilot build kept Cline and Amazon Q Developer in VS Code and exposed models fine-tuned from Qwen3 32B as MCP endpoints for method and unit-test generation. The post says AUMOVIO is gradually migrating its projects to Kiro.
- Vendor case study (AWS for Industries blog, co-written with AUMOVIO)
Deployment details
Automotive embedded software. A hackathon, then a pilot build with AWS, using Cline in VS Code.
- Vendor case study (AWS for Industries blog, co-written with AUMOVIO)
- Team
- AUMOVIO high-performance computer software teams, with the AWS Generative AI Innovation Center
- Workload
- Bug fixing, refactoring, and code explanation; later method and unit-test generation from system designs
- Model
- Claude 3.7 Sonnet in Amazon Bedrock; fine-tuned Qwen3 32B behind MCP in the pilot build
