Explainable daily-routine baselines from timestamped home events.
Give it events such as motion in the kitchen at 08:07 or front door opened at 09:40. It extracts a few daily markers, learns a robust per-home baseline, and explains today in plain words:
No kitchen activity yet by 11:00 (usually 08:10, give or take 25 min).
3 night-time trips between 23:00 and 05:00 (usually 0 to 1).
No movement for 3 h 40 min since 13:20 (longest usual quiet spell about 1 h 50 min).
Built for ageing-in-place and care-taking tools where false alarms erode trust and every alert has to say why. No dependencies, no machine learning, no audio or video: only timestamps and zones.
pip install ./packages/dayrhythm # from this repositoryfrom datetime import date, datetime
from dayrhythm import Event, extract, extract_days, learn, assess
history = [Event(datetime(2026, 9, 1, 7, 46), "hallway"), Event(datetime(2026, 9, 1, 8, 3), "kitchen"), ...]
baseline = learn(extract_days(history, days=[date(2026, 9, d) for d in range(1, 22)]))
now = datetime(2026, 9, 22, 11, 0)
today = extract(todays_events, now.date(), now=now) # an unfinished day
result = assess(today, baseline, now_min=11 * 60)
result.severity # "ok" | "info" | "watch" | "concern" | "learning"
result.summary # the most important explanation, in words
result.findings # every deviation with its z-score, observed and usual values
result.markers # today vs usual for each marker, ready to chart| Marker | Meaning | Window |
|---|---|---|
wake |
first movement of the morning ("morning movement") | 04:00 to 14:00 |
first_kitchen |
first kitchen motion | 04:00 to 14:00 |
first_door |
first front-door opening (contact sensor) | 04:00 to 20:00 |
night_trips |
separate night-time movement episodes (events within 15 min merge) | 23:00 to 05:00 |
quiet_daytime |
longest stretch with no movement (counted up to now for today) | 08:00 to 21:00 |
last_activity |
last movement of the evening | 18:00 to 02:00 |
activity |
number of motion events | 04:00 to 24:00 |
Define your own with FirstEvent, LastEvent, CountEpisodes, CountEvents and LongestGap.
- Baselines use the median and a robust spread (1.4826 x median absolute deviation) so one unusual day barely moves them. Spreads have floors (20 min for times) so a very regular person is not flagged for a small lie-in.
- A marker that is still missing is only flagged if it normally happens on at least 75% of days.
watchstarts at 2 spreads,concernat 3. Quiet spells are never flagged under 90 minutes (watch) or 150 minutes (concern).- Fewer than 7 days of history returns
learninginstead of guesses.
pip install pytest && pytest packages/dayrhythm/testsMIT licensed. Originally written for Hearthbeat, a privacy-first care companion for older adults who live alone.
