MQTT + IoT Core for the edge
MQTT is the standard for lightweight device messaging, and IoT Core terminates it as a managed broker. Cans stay dumb publishers; the cloud handles fan-out, so adding cans costs nothing architecturally.
Project Overview · Computer Vision + Cloud
CarbinWatcher
A computer-vision smart trash can that tells you which bin to use — backed by a serverless AWS pipeline designed to grow from our one hackathon prototype to 1,000+ cans.
- 3rd Place · UCSD DataHacks 2026
- 1,000+ Cans the pipeline is designed for (demoed on one)
- Real time MQTT stream to a live dashboard
The problem
Most people sort waste wrong, and one trash can can't scale.
Recycling only works when items land in the right bin, but the rules are confusing and people guess. CarbinWatcher puts a camera on the can: it classifies each item with computer vision (YOLO + OpenCV) and tells the user which bin to use.
The hard part isn't one smart can, it's many. A real deployment means a fleet of cans all streaming images and classifications at once, a dashboard showing them live, and an analytics layer turning the raw stream into environmental impact. That has to hold up whether there's one can or a thousand, which made this an architecture problem as much as an ML one.
Architecture
A serverless streaming pipeline that scales with the fleet.
Each can publishes over MQTT into AWS IoT Core, which fans the data out two ways: a real-time path to the live app, and a durable path for analytics. Because every stage is managed and event-driven, the same design serves one can or a thousand with no servers to babysit.
Key decisions
Why it's built this way.
Split hot and cold paths
DynamoDB serves the live app with low-latency reads of the latest classifications, while Kinesis → S3 keeps a durable, cheap record for batch analytics. Each path is tuned for its job instead of forcing one store to do both.
S3 → Databricks for analytics
Linking S3 to Databricks let me run scheduled XGBoost jobs over the accumulated stream to estimate environmental impact, without coupling analytics to the real-time serving path.
Serverless-first under a deadline
Every managed, event-driven stage meant no servers to provision during a hackathon, and a design that already scales. The demo ran on the same architecture a fleet would.
What shipped
A working demo, live at DataHacks.
The full loop worked end to end: point the camera at an item, watch the classification appear in the live dashboard within moments, streamed through the real pipeline. Here's the walkthrough.
Challenges & what's next
Honest edges.
Wiring a multi-service AWS pipeline correctly under a hackathon clock, IAM, IoT rules, and the Kinesis → S3 handoff, was the real work; getting the permissions and message routing right took more time than the model. Next would be hardening the fleet story: per-device auth, an on-device model to cut bandwidth, and a proper backfill for the analytics layer.
Stack