Why this page is simulated
The original depended on its trained workflow, local camera, and edge inference environment. Replaying the control logic is more honest than substituting an unrelated browser classifier.
Project 02
CleanSort-CV
An edge vision application that classifies waste from a live camera, tracks category counts, and turns detections into simple operational alerts.
- Completed
- 2025
- Interface
- Streamlit
- Inference
- Roboflow Workflow
- Runtime
- Edge camera system
Interactive walkthrough
Follow a detection through the system.
This browser version replays the application logic without claiming to run the original trained workflow or edge model.
Live camera input
OpenCV reads the camera stream while a separate inference pipeline processes frames. The UI keeps displaying the newest available result instead of blocking on every frame.
Five-class detection
The workflow maps detections to recyclable, kitchen, other, hazardous, and glass waste. The callback receives annotated frames and class names.
Replay the counting logic
Select an example category, then process it. This recreates the state changes from the original Python callback, not model inference.
Operational output
Counts update twice per second. When a configured bin reaches capacity, the application surfaces an alert long enough for an operator to act.
Bin full. Empty now.
System design
Built around the deployment, not just the model.
- Camera threadContinuously supplies the latest frame.
- Inference callbackReturns annotations and class names.
- Thread-safe stateShares the latest frame and counts.
- Streamlit UIRenders video, totals, and alerts.
Interview project
I built CleanSort as part of my Roboflow interview process. The project demonstrated that I could connect computer vision to a usable field workflow, and it helped secure the role.
What I would add next
A recorded waste stream with saved predictions would provide a deterministic demo while preserving the behavior of the original model.
Read the source ↗