I’m building a reinforcement-learning framework that lets industrial sensor nodes tune their own AI workload on the fly so they never waste a milli-joule yet still hit the performance targets we set. Everything runs at the edge—no cloud fallback—so the agent must observe current power draw, latency, and accuracy in real time and then decide whether to prune layers, change quantization, or switch to a lighter model variant. What I need from you • A modular Python implementation that couples a lightweight RL agent (e.g. PPO, DQN or similar) with model-scaling actions exposed through TensorFlow-Lite, PyTorch-Mobile or ONNX Runtime. • A realistic testbed: either an emulator or a Raspberry Pi/Jetson-Nano class board fed with simulated industrial sensor streams plus an energy-measurement interface (INA219, PowerAPI, or your preferred tool). • Training scripts and clear evaluation notebooks that prove the agent meets three objectives—energy efficiency, performance optimisation, and model-scalability—against a static baseline. • Concise documentation so another engineer can port the solution to proprietary sensor hardware later. Acceptance criteria 1. Demonstration video or live session showing the agent running entirely on the edge device, adapting model size/configuration in response to synthetic load shifts. 2. Metrics report that evidences at least a 25 % energy saving while maintaining ≥95 % of baseline inference accuracy and ≤10 % added latency. 3. Clean, well-commented code and a README with setup steps that reproduce your results. If you’re comfortable marrying reinforcement learning with on-device AI optimisation, let’s discuss your approach and timeline.