What it changed
- Weather prediction models built and trained on years of the company's own field data.
- GPU training on demand on Amazon SageMaker: no hardware to buy, and nothing running between experiments.
- Forecasts that turn a sensor network into a predictive product.
The situation
A sensor network tells you what the weather is doing now, in the exact place a farmer cares about. What a farmer needs to know is what it will do next: when to spray, irrigate, plant or harvest. The company had something rare: years of hyper-local readings from its own devices. The question was whether that data could predict the weather better than generic forecasts for those exact locations.
What I did
I built machine-learning algorithms and neural networks, up to deep neural networks, to forecast the weather from that data. That meant the full data-science cycle: collecting and cleaning years of readings, engineering the features that carry signal, training and comparing models, and validating them honestly against what actually happened.
Training deep networks needs GPUs, and buying them for a few weeks of experiments makes no sense. Amazon SageMaker was the right tool: spin up GPU instances to train, pay for the hours used, shut them down, and keep every experiment and model versioned.
Why it matters to you
This is AI in the old-fashioned, hard sense: not a chatbot on top of someone else's model, but models designed and trained on a company's own data to predict something valuable. My master's degree specialised in machine learning, and this is the kind of problem I love. If your business sits on data nobody has asked the right question of yet, that's where to start.