CleanCycle — recycling sorting on Bee

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for CleanCycle — recycling sorting on Bee

Inspiration

For urban households, contamination rates are high because the rules differ by council. We wanted something that just handled it instead of adding another dashboard. Bee was the obvious place to put it because always-on wearable capture is already how this audience works.

What it does

CleanCycle takes sorting household waste correctly end to end. The planning step builds a concrete plan; a memory step remembers what worked and what did not, so repeat suggestions stop; a scheduled job re-runs the plan on a cadence and pushes the output to where it is needed. Everything is reachable through always-on wearable capture.

How we built it

Python plus LangGraph for the agent, Amazon Bedrock for inference, AWS Lambda for runtime, Aurora Serverless v2 for state. We deliberately kept the tool graph small — fewer tools, fewer failure modes — and leaned on Step Functions for everything that has to happen on a timer rather than on request.

Challenges we ran into

Keeping the plan coherent when memory changed the constraints mid-decision was the hard part — an early build happily produced plans that contradicted themselves. Cold starts were the other issue: anything on the always-on wearable capture path has to feel instant, which pushed us to pre-warm the Amazon Bedrock client and shrink the tool graph.

Accomplishments that we're proud of

The feedback loop closed properly: the assistant adapts rather than repeating itself, and urban households noticed within a few days. Getting a LangGraph tool graph to stay coherent under real state was the part we were least sure about and it held up.

What we learned

Agent graphs are cheap to build and expensive to debug — structured traces were the single highest-return thing we added. And designing the evaluation rubric first would have saved us a week of arguing about vibes.

What's next for CleanCycle

Next we want shared use so two people can edit the same plan, a cost estimate driven by live pricing, and a second surface — moving from a wearable to a companion view so the same agent is usable in more than one room.

Built With

• amazon-bedrock
• amazon-bee
• aws-lambda
• cleancycle
• langgraph
• waste