Try it
Bring your own delivery data
Our exact pipeline runs in your browser. Nothing is uploaded.
Load a LaDe-format delivery file, optionally poison it with the same defects we studied, and watch a gradient-boosted model and the Guard train right here. The Guard's settings were frozen on Chongqing, so whatever you see on another city is an honest out-of-sample test, including when the Guard doesn't help.
Python Starting Python in the background…
Demo not working? Watch this video of a live run instead.
A 40-second screen recording of this page running the synthetic sample (Sloppy couriers 20%), from loading the file to the results.
1 Load data
8 weeks, 45 couriers, about 23,000 packages. Synthetic: generated by scripts/make_sample.py, not real deliveries.
For the demo, load the real LaDe Yantai file delivery_yt.csv (31 MB, 206,000 packages) from the LaDe dataset. It never leaves this tab.
LaDe times have no year (MM-DD HH:MM:SS); the year sets the weekdays.
2 Check data
Load a file to see its audit: format, weeks covered and the issues already in it.
Issues already in the file
Check (same as src/audit.py) | Rows | Share |
|---|
3 Poison it (optional)
Pick what goes wrong. Defects are injected into the training weeks and the history features only; the test week is always scored on the durations in your file.
4 Train
5 Results cached, not live
Run the experiment to see the results.
Error per run
Detector scorecard
Positive = a training record the poison dial changed or created. The injector's labels are used only here, to score the Guard; the Guard never sees them.
Dispatcher review
Dispatcher time is an assumption: 45 seconds to look at each deferred package.
Downloads
Made in your browser from this run.