Robotinc

Part 8: What We’ve Learned & Teardown

What We Built

Did we over-engineer a light switch? Absolutely. But along the way:

Hardware: Dusted off an M5Stick, discovered most USB-C cables are lies (power-only), and got an ESP32 to listen to MQTT commands.

Security: Moved from a public MQTT broker (hello, internet strangers) to AWS IoT Core with X.509 certificates. Because authentication matters, even for LED screens.

AI Layer: Built a multi-agent system with Bedrock Nova Lite and Strands. The agent understands “turn it ocean blue” and doesn’t hallucinate RGB values, thanks to tools and the Python color library.

Edge Deployment: Automated everything with CDK. One cdk deploy creates VPC, EC2, Greengrass, DynamoDB, Lambda, IoT Rules, and pushes a Docker image. Infrastructure as code is satisfying when it works first time.

The Result: Natural language controls physical hardware. “Make it sunset orange” → agent thinks → MQTT publish → screen changes. Physical AI in action.

Key Lessons:

  1. Physical AI is different: Tools prevent hallucination, state must be verified
  2. Edge + Cloud works: Low latency at the edge, AI inference in the cloud
  3. Infrastructure as Code: CDK makes deployments repeatable
  4. Hybrid approach: Greengrass for connectivity, Docker for flexibility

You now have the foundation to build Physical AI applications that bridge digital and physical worlds.

Teardown

To avoid ongoing charges:

cd infrastructure/cdk
cdk destroy

# Confirm when prompted
# Removes: EC2, VPC, DynamoDB, Lambda, ECR, IoT Thing, certificates

Cost Reality: ~£10/month if left running (t4g.small + Bedrock usage). Always destroy when done.

Next Steps

Consider extending this project:

Resources