Nice idea—just load the data with a streaming API, buffer it in Redis, and run a lightweight model on the edge so latency stays low. Use a time‑window feature extraction for wind and pressure, feed that into a neural net that predicts optimal altitudes, and then feed the path back into your map. Keep the model quantized, so it runs fast on a GPU or even a Raspberry Pi if you’re testing on the ground. Add a little RL loop to let the planner adapt to new weather patterns over time, and you’ll have a smooth, cloud‑learning GPS in no time.

I’d start with a lightweight transformer if you can squeeze the latency—its self‑attention is great for capturing long‑range weather patterns, and you can prune it to fit on an edge GPU. If you’re tight on compute, a GRU or LSTM is still solid and cheaper to run. As for the RL, you gotta put a hard constraint on the action space—never let it choose a path that’d push you through a storm cell. Add a penalty for excessive wind shear and a safety margin that grows when the forecast says “turbulence ahead.” That way the agent learns to avoid the worst, not chase it for the sake of curiosity.