Why I built Upgrade Pro.
A calibrated answer to a question the airlines won't answer: am I actually going to clear?
I spend a serious share of my life in seat 12C, hoping it turns into 2A. I also spent years building machine-learning models for a living. For a long time those two facts had nothing to do with each other.
What changed is that I got tired of guessing. You book a flight, you clear or you don't, and nobody tells you why. So you collect folk wisdom instead: fly Tuesdays, avoid the Monday morning bank, the 6am is always wide open. Some of it is true. Most of it is someone's small sample dressed up as a rule. Meanwhile the thing that actually decides your seat — how many premium seats are open, how many elites are ahead of you, what the fare buckets look like four days out — is sitting right there in the data, and no tool would just tell me the odds.
So I built one for my own trips. Not an alert that a seat opened, and not a promise I'd clear: a probability, with an honest range around it, for every departure on the route. The first time it told me my 8am was a long shot and the 4:40pm was a real contender, I moved my booking and cleared. That was the whole product in one afternoon.
Sharing it felt obvious. The rule I gave myself is the one thing I'd want from someone else's model: never dress up a guess. When a number is estimated, we label it estimated. When a flight is hopeless, we say so before you book, even though "don't bother" is a strange thing for a product to tell you. And every prediction gets graded against what actually happened at the gate, in public, including the misses.
What the numbers say.
Upgrade Pro reads the signals a gate agent sees — standby and upgrade waitlists, fare-bucket inventory, seat maps, aircraft type — and feeds them to a random-forest model, isotonically calibrated on tens of thousands of completed flights. On graded predictions it scores 0.91 ROC-AUC, meaning that given one flyer who cleared and one who didn't, it ranks the right one higher about 91% of the time; a model with no idea scores 0.50. Its Brier score is 0.112 — roughly half the error you'd get by guessing the historical average for every flight. Those figures move as the model retrains, and the current ones are always on the track record page, misses included.
Tell me when it's wrong.
If a prediction looks off, or the product does something stupid, I want to hear it — that feedback genuinely shapes what gets built next. Email we.are.upgradepro@gmail.com.
Upgrade Pro is operated by Third Culture LLC.
See what it says about your next flight.
Your first search needs no account and shows every flight on the route, scored.