This project focuses on the complete lifecycle management of Machine Learning models (TinyML) deployed across resource-constrained Internet of Things (IoT) devices. While training models at the cloud or edge is increasingly common, managing their deployment, monitoring performance, handling incremental versioning, and executing Firmware Updates Over-The-Air (FUOTA) under strict network limitations (such as LoRaWAN constraints) remains a major architectural challenge. Students will design and implement an end-to-end management platform that bridges cloud-based MLOps pipelines with constrained IoT hardware fleets.
Focus
Designing, implementing, and evaluating a holistic Cloud-to-IoT MLOps and management platform. Emphasis on automated model versioning, delta-based updates over resource-constrained networks, fleet-wide monitoring, and orchestrating the entire TinyML lifecycle from cloud training to edge execution.
Tasks
- Platform Architecture: Design a centralized cloud/edge management service that coordinates model training, version control, and packaging for resource-constrained targets.
- Efficient Update Mechanism: Implement or integrate delta-based update strategies (e.g., incremental model weights) optimized for constrained transmission protocols like LoRaWAN FUOTA.
- Fleet Monitoring & Observability: Build telemetry and monitoring loops to track device-level resource constraints, inference performance, and model drift across distributed IoT fleets.
- Evaluation & Benchmarking: Deploy the platform across a testbed combining cloud services and physical or simulated constrained nodes, quantifying update delivery times, bandwidth consumption, and resource overhead.
Required Skills
Theoretical Skills:
- MLOps and Machine Learning Lifecycle Management
- Constrained Device Computing & TinyML Concepts
- IoT Networking Protocols (LoRaWAN / FUOTA)
Practical Skills:
- Full-stack development (Python, C/C++ for microcontrollers)
- Cloud deployment and containerization (Docker, Kubernetes)
- Experience with embedded ML frameworks (e.g., TensorFlow Lite for Microcontrollers)
