Weathered

Weathered started as a final-year project with an awkward brief: do something ambitious with public climate data, on a student budget, without spinning up a server.

The Met Office publishes a century of monthly readings across dozens of UK stations. The raw files are small, but the cross-station analyses I wanted to do — long-horizon anomalies, region-vs-region comparisons — needed a pipeline that could fan out and come back cheap.

I landed on a Lambda-driven ETL: scheduled fetchers write normalised Parquet to S3; a second Lambda pushes feature tables into SageMaker for anomaly detection; a thin Next.js frontend reads a pre-rendered JSON manifest and hands chart specs to Plotly.

The frontend is deliberately quiet. No filter sidebars, no dropdowns stacked four deep — each view is a single chart and a short paragraph. If a user wants a different cut, the URL changes; the UI doesn't grow.

Shipped on time, marked well, and more importantly taught me how much product thinking survives even the narrowest academic brief.

The pipeline runs under the AWS free tier. I still use it to pull temperature anomalies when an article makes a claim I want to check.