Zane Bergen

Sleep apps often reduce your sleep to a standardized score broadly applied to all users. The number I care about is directly related to how rested I feel, and that relationship is different for each person. None of the apps I used were fitting a model to my individual sleep patterns and how rested I felt each day.

1. Built an iOS app in Swift and SwiftUI that pulls each night from the Oura API and pairs it with a one to five rating the next morning. The rating is the label and the biometrics are the features.
2. Fit an n-of-1 model that refits itself every night, entirely on the device. The data lives in a local GRDB database and never leaves the phone.
3. Designed the whole thing around honest uncertainty. Last night's data stays locked until you rate it, and no weight is shown until it clears a confidence threshold.

A working build that gets progressively more defined each morning. Early on the honest answer is usually "too uncertain to say", and the app is allowed to say it. The screens below walk through one morning.

Garmin and Strava log a run as one data point. A ten mile day with six kilometer repeats inside it looks the same as an easy ten. Effort needed to be added to the equation. I wanted to build a dataset of my 12-year running history, accurately labeling all 3,370 individual workouts by type.

1. Parsed every FIT file into a uniform one-second series and split each run wherever the pace shifted, using PELT change-point detection. That produced 32,213 segments.
2. Clustered the segments into seven effort states with k-means. Every segment is scaled against a trailing 90 day window, so hard means hard at that month's fitness rather than my all-time best.
3. Hand labeled 272 runs blind, wrote a transparent nine-rule baseline, then trained a gradient-boosted model on features describing how each run is composed.

The gradient-boosted model reaches 79.7% strict accuracy on out-of-fold test runs, 22 points above the rules baseline. The labeled corpus, 2,969 runs that are roughly 70% easy and 20% quality work, is the foundation for the next stage, predicting race times from training history.

When a crash closes a road, neither navigation apps nor traffic agencies have a good estimate of how long it will stay closed. We asked whether public accident data could predict clearance time well enough to route around it.

1. Drew a stratified sample of 100,000 crashes from 7.7 million U.S. accidents (2016 to 2023), engineered clearance time as the target, and removed every column that leaked the answer.
2. Cross-validated five regressors, from an OLS baseline and ridge through random forest and two flavors of k-nearest neighbors, one of them with TF-IDF text features from the crash descriptions.
3. Ran inference on the linear model with a Benjamini-Hochberg correction to see which predictors actually held up on the test set. Only six numeric features and two road indicators survived.

With Nick Narey and Trevor Moore. I owned the OLS baseline, the ridge pipeline, and the statistical inference.