

An alarm that refuses the snooze you can’t afford
Setting an alarm took me a long time before bed, and half awake I often judged wrong and overslept. So I built one that stops the snooze at your last safe minute, then tested it against just getting up early.
Role Product, model evaluation, the backtest, the replay and the videoResult Best Use of TabPFN in Week 3 of DEV’s Hacktoberfest Open-Source AI Challenge
Does it beat just getting up earlier?
Overall, no. A 13-minute buffer, between 10 and 15 on this slider, was late 5.1% of the time, the same as the alarm, at about the same wake-up time. But I only found that number by looking at August, and you can’t do that the night before.
The alarm helps most on hard routes. On the University Heights route it got you up about 7 minutes earlier than that 13-minute buffer and was late 12.7% of the time, against 17.1%.
Move the buffer and compare each route with the alarm.
- Late, all routes
- 8.3%alarm 5.1%
- Late, worst route
- 22.9%alarm 12.7%
- Minutes early at arrival (median)
- 11.4alarm 13.6
Busy routes
- Greenpoint → Midtown East
- Hunters Point → Midtown Center
- Upper East Side → Hudson Yards
- Battery Park City → Hudson Yards
- TriBeCa (Center) → Midtown Center
- TriBeCa (East) → Midtown East
Routes with few trips
- University Heights → Midtown Center
- Cobble Hill → Midtown North
- East Harlem → Financial District
- Heartland Village → Financial District
- Melrose → Midtown East
- Queens Plaza → Midtown North
The usual rule, and here the alarm wins. Move to 15 and it no longer does overall.
1,070 mornings on 12 routes in August 2026, scored on real Uber trips. Routes with few trips borrow rides from any August weekday, so they show a typical weekday, not that day’s traffic. The University Heights figures cover 63 mornings with meetings at 8:00, 9:30 and 10:00. Route figures are rounded.
Try the demo morning
This is a 9:00 meeting, University Heights to Midtown Center. The alarm rings at 07:14 and you can snooze until 07:44, when the chance of being on time is 90%. One minute later it is 89%, so the snooze stops.
The replay runs from results I computed in advance, so no model runs in your browser.
A probability for every wake-up time
The night before, TabPFN-3.5 predicts the range of ride times from up to 10,000 similar July trips. For each possible wake-up time I simulate 2,000 mornings, each with about 15 minutes to get ready and then the ride. The cutoff is the latest time with at least 90% of them on time.
The usual 10-minute rule on this morning means getting up at 07:57, with a 51% chance.
What did not work
At airport pickups the plain lookup table beat TabPFN, late 9.1% of the time against 10.3%, so I would improve those predictions before trusting it for a real pickup.
On the University Heights route the alarm is still above its own 10% target.
I also dropped a feature. An early idea was “wait 10 minutes to save on the fare”, but between neighbouring 10-minute slots the median fare moved by only about $1–2. Fares inside one slot varied 10–15 times more.
- CONTEXT
- Hacktoberfest challenge entry, Oct 2026. A web replay and a command-line planner, not a phone app. The phone in the opening picture is a rendering. It doesn’t know about tomorrow’s accident or weather.
- DATA
- NYC TLC trip records and BTS on-time data; July to learn from, August to test. Getting ready (15 minutes) is an assumption.
- STACK
- TabPFN-3.5 · LightGBM · Python · a static replay page · built with AI coding agents
The replay, the video and the code
The code is Apache-2.0. TabPFN-3.5's weights are free under Prior Labs' non-commercial license, and the repo links every report behind these numbers.