Every time a geomagnetic storm makes the news, the same thing happens in my household: someone asks "can we see the northern lights tonight?", and I open three tabs — a Kp index chart, a cloud forecast, and a light pollution map — and still can't give a straight answer.
The existing aurora tools are built by and for space-weather people. They show you the raw data: planetary K-index, aurora oval plots, solar wind speed. That's great if you know what "Bz = -12 nT" means for your driveway. Most people don't. Some sites go the other way and print invented precision like "73% chance of aurora tonight" — a number nobody can honestly produce.
So I built Aurora Tonight (aurora-tonight.com): enter your city or ZIP, get one of four answers — GO, MAYBE, NO, or UNKNOWN — plus the best viewing window, the main obstacle in plain language, and which direction to look.
The stack
Nothing exotic, and that was the point:
- Next.js on Vercel. Forecast pages are served via ISR — the answer for a given city only changes as fast as the upstream data does, so per-request rendering was just burning invocations for identical bytes.
- NOAA's aurora oval data for geomagnetic activity. This is the authoritative public source; I deliberately don't layer a proprietary "prediction" on top of it.
- Open-Meteo for cloud cover snapshots. Clouds are the #1 reason a strong geomagnetic night is still a NO, and most aurora sites ignore them entirely.
- Darkness, moonlight, and light pollution are computed or looked up server-side. Everything is cached at the edge; the whole site is static-ish and free to run, which is why the tool can be free with no signup.
The interesting design decision: what not to compute
The hardest part wasn't the data pipeline — it was deciding the output contract. Three rules I ended up with:
- No fake percentages. The model's output is a categorical call, not a probability. If you can't defend "73%", don't print it. GO / MAYBE / NO maps to what a person actually does with the answer.
- UNKNOWN is a first-class state. If the cached data snapshot is stale, the page says UNKNOWN instead of extrapolating a confident-looking guess from old data. This costs us some "always has an answer" slickness, and it's worth it.
- The main obstacle is part of the answer. "NO" alone is useless. "NO — overcast until after 1am" tells you whether to set an alarm.
I wrote up the full scoring logic on the methodology page if you're curious how the calls are made — it documents exactly what goes into a GO versus a MAYBE.
Current scope and honest limitations
It covers 15+ indexed US states and cities — Fairbanks and Alaska down through mid-latitude places like Colorado, Ohio, and Chicago, where aurora is a rare-event question rather than a nightly one. Northern hemisphere, mid-to-high latitudes only. It won't tell you anything useful in Texas, and it says so.
If you're building anything in the "raw public data → plain-language decision" space, I'd genuinely like to hear how you handled the honesty-vs-slickness tradeoff. The temptation to print a number that looks more confident than the data deserves is real.
