Purrtector

Dashboard

Built phone-first: one column, and every indicator opens its own trend in place.

• Calendar — one mark per day from the day's status: asterisk for normal, a smaller one for a warning, a red cross for danger.
• Indicator sheet — the selected day's eight values, each as a bar against the 14-day band. Tap a row and its trend opens under it, over 2 weeks, 1 month or 3 months.
• Alerts — behind the bell, ordered by severity.
• Log — every visit of the day, with a tag control that feeds the urine/stool model.

Circuit

Load cells under the box, an ESP32 that detects visits, and a web service the phone reads.

• Sensing — four 50 kg load cells into an HX711.
• ESP32 — detects each visit and logs it.
• Web service — stores events and serves the dashboard to the phone.
• Tail gadget — a tail that rises when he is well and drops when something is off, so his health reads at a glance without opening the dashboard. Still being built.

Algorithm

One weight signal is turned into events on the device, and events into daily judgements in the browser.

1. Visit detection (C++, on device) — a state machine over the smoothed weight. Over 1.5 kg for 1 s means a cat is in; under 0.7 kg for 2 s means it left; the box must then hold within 15 g for 5 s. The settled difference is the deposit, the mean load while inside is the cat's weight. Visits under 5 s are dropped.
2. Cleaning and drift — a change over 20 g with no cat is logged as cleaning, as is a press that leaves the box lighter. Idle drift is tracked with an EMA (α = 0.02). A “visit” over 15 minutes is an object on the scale and triggers a re-baseline.
3. Classification — deposit under 10 g: a failed attempt (or a hop-in if under 30 s). Urine and stool weigh alike, so visits I tag train a nearest-centroid model on log mass and log duration; after five of each it labels the rest as estimates.
4. Personal baseline — eight daily indicators, each judged against the mean ± spread of the cat's own previous 14 days (defaults until 3 days exist; count bands rounded to whole numbers).
5. Alerts — three failed attempts within 6 h is danger; any failure, 1.5× the usual visits, or 12 h without output is a warning.

Stack