Hao Sun

About 72% of buses fall inside the official window: no more than 1 minute early or 5 minutes late. That window makes early buses look more common than late ones. With the same 3-minute cut-off on both sides, buses are late about four times as often as they are early.

The feed only gives predictions, never actual arrivals, so I take the last prediction made within 150 seconds of the stop as the arrival time. Raw snapshots are kept untouched, so the whole dataset can be rebuilt if that logic changes. A quality report checks for collection gaps and how close to arrival the predictions were made (median 42 seconds).

Figures are provisional while collection continues.

Models are layered: staging views clean and rename the raw tables, intermediate models join orders to line items and calculate discounted amounts, and a mart model produces fct_orders, ready for BI. dbt tests, including referential checks, run on every build.

Cosmos runs each dbt model as its own Airflow task, so a failure shows exactly where it happened. It connects to Snowflake with key-pair authentication and runs locally in Docker through the Astronomer CLI.

Insights has a daily completion chart with a 7-day rolling mean, a habit-by-day grid, and phi correlations between each pair of habits over 60 days. The mood view estimates how much higher you rate your mood on days you did each habit, with 95% confidence intervals, and plots completion against mood with Pearson’s r.

Every chart has a table view, and the data exports to CSV with one row per habit per day, ready to load into pandas. Each push runs the tests and deploys to GitHub Pages.