CarbonClever: Engineering Real-Time Carbon Accounting
Full-Stack System · React.js, FastAPI, Docker, K8s · 2024-2025
Figure 1: Visualization of complex user trajectories processed by our system.
Bridging Research and Reality
Academic models often live in clean, static datasets. I wanted to see if our Speed Transformer model (see Project 1) could survive contact with the real world.
I led a 7-student team to build CarbonClever, a WeChat mini-program designed to track individual carbon footprints in real-time. This wasn't just a prototype; it was a production-grade system handling data from 348 real users across China.
System Architecture
The core challenge was processing high-frequency GPS data from thousands of heterogeneous devices (Android/iOS) with varying signal quality.
- Frontend: Built with React.js within the WeChat ecosystem for maximum reach. I designed an item-based recommendation system that drove in-app engagement up by 50%.
- Backend Inference: Exposed the PyTorch-based Transformer model via FastAPI and Docker containers, enabling real-time mode detection ( < 500ms latency).
- Data Pipeline: Engineered a Dask-based distributed pipeline capable of transforming 4TB of raw GPS logs into training-ready features in under 20 minutes.
- Infrastructure: Deployed on AWS EC2 with Kubernetes and ArgoCD for automated CI/CD, maintaining 99.9% uptime during the field experiment.
Impact
4 TB+
Of GPS data processed efficiently.
150%
Achieved 150% of our recruitment target through targeted social ads and data compliance trust.