Flight Finder.

The subscription services that email you cheap flights work off a known mechanism: cached fares from real searches, odd routings, prices that briefly fall through their own baseline. I’d tried the idea once before as a custom GPT for reward flights — used it once, shelved it. The idea was right; the form was wrong. It needed to be a pipeline, not a conversation.

So: discovery pulls candidate fares from cached market data across ninety-plus routes out of London. A verification pass confirms the interesting ones against live fare searches before anything reaches me. Every candidate gets an explainable 0–100 deal score across nine weighted components — discount against baseline, destination preference, stop quality, trip-length fit, departure time, absolute price. Nothing scores against a guess: baselines come from the system’s own recorded price history once it’s warm, market medians before that, and every alert stores the baseline and method behind its claimed saving.

Alerts land three mornings a week — extraordinary deals fire immediately, the rest arrive as a digest, capped and cooled down so Barcelona can’t spam the feed. Each one carries tap-to-book links, and a follow-up push brings the destination weather: a real forecast if the trip is close, the same dates last year labelled honestly as “typical” if it isn’t. Long-haul deals only qualify if the stay justifies the flight — no £1,000 weekend in Sydney.

The whole system costs nothing to run. No database, no server — a scheduled workflow verifies, scores, rebuilds the dashboard, and commits its own state back to git three times a week. The frontend is one hand-written page: a draggable globe with great-circle arcs to every alerted destination, and an explain popover on every price that shows exactly why it mattered.

→ flight-finder-green.vercel.app