Inspiration

We were inspired by a real-world gap: vehicle owners often report the same safety-related issues independently, but manufacturers and dealers usually discover the pattern only after many cases have already occurred. We wanted to explore whether an agentic AI system could detect these early warning signals from everyday conversations while still keeping a human in control of any customer-facing action.

What it does

Loopback Early Warning collects vehicle issue reports, structures them into cases, and automatically detects clusters of similar symptoms. When a pattern emerges, it alerts the engineering team. Only after human approval does the system notify the affected owners with available dealer appointment slots. The agent follows a strict safety-first policy and never provides a diagnosis or says a vehicle is safe to drive.

How we built it

We built the solution as a serverless application on AWS using Amazon Bedrock AgentCore for the agent, AWS Lambda for processing logic, DynamoDB (with Streams) for case storage and clustering, and Amazon SES for alerts and notifications. Infrastructure was defined with AWS CDK. WhatsApp and RCS were designed as primary messaging channels, with SES as the currently working notification path.

Challenges we ran into

Integrating multiple AWS services and making the agent behave reliably was challenging. Designing a prompt that stays helpful without crossing into diagnosis required careful iteration. We also had to balance the full multi-channel vision with what could be completed within the hackathon timeline, while ensuring the system remained responsible through a human approval step.

Accomplishments that we're proud of

We successfully created an end-to-end agentic flow that includes safety guardrails, structured case handling, automatic clustering, and a clear human-in-the-loop control point. We are proud that the system prioritizes safety and responsibility rather than full automation.

What we learned

We learned that building useful agentic systems requires strong tool design, clear safety policies, and thoughtful human oversight. We also gained practical experience with Bedrock AgentCore, event-driven architectures using DynamoDB Streams, and the real-world constraints of messaging channels.

What's next for Loopback

Next we plan to fully integrate WhatsApp and RCS as primary channels, improve clustering accuracy, add richer dealer booking flows, and explore partnerships with fleet operators or automotive service platforms to test the system with real-world data.