Weather Explorer: backend
The server half of Weather Explorer, a prototype for exploring animated ECMWF weather
forecasts. Once a day an AI research agent picks about fifteen interesting weather situations
around the world ("highlights"), backs each one with evidence, and describes how to show it. This
backend runs that research, prepares the data for each animation, and serves the result as a JSON
feed.
Please read Important caveats before using anything here. This is a
prototype. It is not an ECMWF service, its weather content is model guidance chosen and written
by an AI model, and its code was written by an AI model.
Contents
Who made this, and how
All of the code and documentation in this repository was written by Claude Opus 5.5, an AI
model made by Anthropic, working in
Claude Code at the direction of Neil Gordon. Neil set the
goals, the design rules and the priorities, tested the results and decided what to keep. Claude
wrote the code, ran the tests and wrote the commit messages.
Neil developed Weather Explorer as an ECMWF Visiting Scientist. ECMWF supported
the work by providing the Claude Opus 5.5 usage, through its enterprise account, with which the
code was written. That support does not make it an ECMWF product or service (see below). Commits it made end with a
Co-Authored-By: Claude line.
The rules Neil set are worth knowing because they explain much of the design:
- ECMWF Open Data only. Nothing restricted (no MARS, no Extreme Forecast Index), so anyone can
reproduce every field shown.
- The server only researches and serves highlights. It does no work per viewer. Decoding,
colouring, contouring and animation all happen in the browser or on the phone. (An earlier
version did per-request work on a 0.5 vCPU server and froze.)
- Honest sourcing. Model output is guidance, not a warning. Official warnings belong to
national weather services, and the feed cites them.
- Low running cost. It runs on a small Replit virtual machine, and the daily research sweep costs
a few US dollars in AI model usage.
Important caveats
Not an official forecast or warning. Everything shown is model guidance from ECMWF's
forecasts, chosen and described by an AI model. It can be wrong, out of date or misleading. For
forecasts and warnings, use your national weather service. Nobody should rely on this for safety,
travel, operations or any decision.
Not an ECMWF service. It uses ECMWF's free Open Data and was developed with ECMWF's support
(see above), but it is not an official or operational ECMWF product or service. ECMWF does not
run it, maintain it or vouch for its content.
AI-chosen and AI-written content. The highlights, their summaries, captions and citations are
written by an AI model (currently GPT-6.1 Sol from OpenAI; Claude Opus 5.5 is the alternative).
The pipeline checks the feed's structure, and the instructions require evidence with real links,
but no person reviews each highlight before it is published. Expect occasional errors in names,
places, numbers or timing.
AI-written code. The code was written by an AI model and tested against the cases described
in this README, not audited line by line. It is a working prototype, not production software. It
may contain bugs, security weaknesses or poor practice. There is no warranty (see the
licence) and no support.
Running it costs money. A research sweep calls a paid AI API (Anthropic or OpenAI) with web
search. If you run your own copy, you pay for your own usage, and you are responsible for keeping
your own keys private.
Third-party data and services. The software downloads data from ECMWF, MeteoAlarm, GDACS, the
WMO and national weather services. Each has its own licence and terms (see
Data sources). Respect them, keep the attribution, and
don't hammer their servers: the code fetches only what it needs, and so should any changes.
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
