Every weather app tells you it's 86°F and clear. None of them tell you whether you can pour concrete today. CrewCast does: a job-site weather app for contractors and outdoor trades, live at stormradar.vercel.app, built around one question — is this workable? — and a lot of machinery devoted to being honest about how confident it is.
What it does
The Today screen leads with a job-site score out of 100 and a plain-language briefing: "Currently 86°F and workable. Dry through the afternoon — a good window for most outdoor trades." Under it, a best-work-window line and a per-trade breakdown.
There are 15 trade profiles — concrete, roofing, plumbing, painting, excavation, HVAC, landscaping, pressure washing, windows, electrical, tree work, solar, pest, asphalt and moving — each a small function evaluating current conditions into go / caution / no-go with a reason. Adding a trade is one definition plus one button. A contractor can also opt into a push notification the evening before a no-go day for their trade.
Beyond that: radar with a scrubber and NEXRAD tiles, precipitation nowcast, hurricane tracking from the NHC, NWS alerts with a contractor-focused recommended action per alert type, a printable weather-delay report with per-cause breakdown and official alert history, and a true PDF export rendered server-side.
Everything runs on free, mostly public-domain sources: NWS, NOAA, NHC, NEXRAD via RainViewer, Open-Meteo, USGS and NASA EONET.
The trust pack
The distinguishing feature isn't a forecast. It's the layer that tells you how much to believe one.
Model agreement. The forecast is fetched from GFS, ECMWF and ICON simultaneously and compared. The briefing carries a line like "Models disagree (GFS · ECMWF · ICON): 2 of 3 show rain, 1 of 3 show storms, temp spread 3°." Disagreement caps the displayed confidence — a split forecast can't be High confidence, no matter how nice the average looks.
Tuning this was instructive. The first version flagged all seven days as split, because a 49% versus 52% rain probability counted as disagreement. It isn't. The criterion became a genuine gap — one model above 60% while another sits below 40%, or a temperature spread over 8° in the near term. Florida in August still comes back genuinely split five days out of seven, which is the correct answer.
Station observations. The app pulls the actual current temperature from the nearest NWS station and shows it — "Observed 88° at KDJT · 24m ago" — flagging in amber when the model and the thermometer disagree by 5°F or more.
A self-verification scorecard. Every day the app snapshots its own next-day high and rain call, then grades itself the following day against station observations where the sensors cooperate and model actuals where they don't, keeping a rolling 30. It says which basis it used. It also shows "collecting — N of 3" until it has enough results to say anything, rather than displaying a meaningless percentage on day one.
A lightning all-clear timer with an honest scope note: it runs a 30-minute countdown driven by NWS signals and storm codes, and states plainly that it is not strike data, because no free real-time strike feed exists.
Alongside those: the forecasters' own reasoning pulled from the NWS Area Forecast Discussion, and SPC convective outlooks resolved by a hand-rolled point-in-polygon test against the day-1 through day-3 GeoJSON.
And running through all of it, wording chosen to be defensible. Not "safe to work" but "likely workable." Every score card carries "Guidance, not a safety guarantee — verify on site," and a safety page that says plainly the NWS is authoritative and OSHA wins.
The gotchas
The service worker had never been deployed, so push was broken on every platform. This is the worst bug in the project's history and it was invisible for weeks. Each frontend deploy copied only index.html into the deploy directory. serviceWorker.register('/sw.js') therefore 404'd, silently, and Web Push simply did not work — not just on iOS, everywhere. The fix was adding sw.js, a manifest and three icons, and the durable lesson is that a deploy step which copies a file rather than the directory will eventually ship a broken app that looks fine. (The iOS-specific reality is separate and also real: Safari tabs don't expose push at all, so the app now detects iOS-not-standalone and shows Add to Home Screen instructions instead of a dead toggle.)
A mapped colour sweep will eat your data colours. Restyling the app from navy-and-cyan glassmorphism to industrial charcoal-and-safety-orange meant 141 scripted replacements across the source. It worked — and it also recoloured the Saffir-Simpson hurricane scale, where tropical-storm cyan collided with category-2 orange. Data ramps are not brand palette. Water-blue for humidity and rain bars was kept deliberately for the same reason. If you sweep colours by literal value, semantic colours are collateral damage.
Official alert history needs two endpoints and real marine filtering. Pulling historical NWS alerts for the delay report means hitting both IEM archives: storm-based polygon warnings, where severe thunderstorm, tornado and flash flood warnings actually live, and the zone/county endpoint for watches and advisories. Then merge and dedupe. And filtering marine noise takes more than excluding one phenomenon code — a coastal job site drowns in Small Craft Advisories unless you filter the whole marine set plus the marine zone UGC patterns. Verified against a known event: a Severe Thunderstorm Warning at a specific West Palm Beach location appears exactly once, at the right time, with zero marine leakage.
Background tabs break animation timing. A bottom sheet used requestAnimationFrame to add its open class. rAF is throttled in background tabs, so the sheet became display: flex while staying translated off-screen — visible as a dead grey area. Replaced with a forced reflow and a synchronous class add.
International users were being silently misled. A test from London revealed the app degrading quietly: NWS feeds fail outside the US, so the alerts panel would have shown "✅ No Active Alerts" — which abroad is not merely wrong but dangerous. Now the app detects coverage two ways and, outside the US, shows a notice explaining that forecasts and radar work fine while warnings and push are US-only, and replaces the all-clear with "🌍 Official warnings unavailable here."
Where it stands
Live, rebranded from StormRadar to CrewCast after a trademark collision, with a marketing landing page, seven legal and trust pages, seven SEO landing pages, a sitemap and security headers. The source was refactored out of one enormous HTML file into 26 modules with a build step and a proven byte-exact round trip — the single-file version is now a build output that should never be hand-edited.
Honest caveats: it's imperial units only, real internationalisation is deliberately not built, and monetising it requires clearing three licensing gates — an Open-Meteo paid plan, a RainViewer commercial agreement or a swap to public-domain NEXRAD tiles, and a basemap plan. Knowing where those gates are is itself part of the work.
One thing this post can't hide, since the screenshot above was taken with the theme pinned: the automatic daylight theme has a race where the page background stays dark while the cards flip light. That's the next fix.
More projects built this way are on the projects page.
