PolyMO Digital Pet Only Lives If You Do Without Doomscrolling

Brenpoly’s PolyMO is a palm-sized desk creature with a bright 1.8-inch AMOLED face, a microphone, a speaker, and a capacitive screen, and its health is bound to how long you stay inside the apps that eat an evening. Sit it on a table, pair it over Bluetooth to an Android phone, and the little face blinks, grows hungry, gets bored, and moves through life stages on its own clock. Spend too long on flagged feeds and it falls ill. Leave the illness untreated long enough and it dies for good. That is the whole contract.

The hardware is a Waveshare ESP32-S3 Touch AMOLED 1.8 board with 368 x 448 pixels and a built-in microphone and speaker. Any of the V1 or V2 panel variants will suffice, as the firmware is intelligent enough to detect which driver you’re using on the fly. 3D printable shells are available in the repository, allowing you to simply put your board into a nice little body to make it appear less like a development kit and more like a finished product. All of the simulation logic for hunger, happiness, age, sickness, and death is executed on the ESP32, which is saved in flash memory and timed using the board’s internal clock. Pull the power cable, and PolyMO will continue to tick without saying anything.

To communicate with PolyMO, simply speak toward the case, and the board will stream Opus audio via BLE. The Android companion app handles the rest, transcribes the footage with Whisper, transmits it to a local language model for analysis, and finally provides the response back via radio. There is no chat window on your phone for the conversation to take place in; instead, you complete all of your setup, select your models, and monitor screen time from there. The models aren’t even provided with the software; you have to extract them from the file system using the system file picker. The program isn’t even asking for an INTERNET permission; if that happens, the build would fail the unit tests. That’s sort of the point of having a device that can hear you and read notifications.

Screen time is more of a variable that changes depending on the weather (which, of course, you have no influence over) than a setting that you can turn on or off. You have a basic allocation for each sitting, which is set to roughly five minutes on the applications you’ve identified as favorites, and you may adjust it in small increments. Crossing the line causes the pet to become ill, not death. Satiety and happiness scores (all of which are integer values ranging from 0 to 4) will decrease faster when the pet is sick. Simply close the app that was making it sick or give it a slap (aka, tap the face), and the illness will be gone. Leave both scores at zero for long enough (approximately 12 hours, unless you muck things up and speed up the process), and death becomes a state that must be willingly restored. Quiet hours from 9 p.m. to 9 a.m. will pause the death timer and most calls, but late-night scrolling may still make your pet sick.

Care is an action-based game; a few touches on the screen or a brisk shake of the case for around three seconds will initiate play. A simple tap on the face can cure a lot of ailments. Simply ignore a call when your score reaches zero and the firmware believes you committed a mistake, as this counts against your lifespan. Life stages progress from egg to kid to teen to adult around every 12, 36, and 72 hours, after which older stages begin to degrade much faster. Facial expressions and personas are all derived from JSON files, so a joyful face must smile, while a stiff face lacks a mouth. Your responses are concise, employ sentence case, and are all influenced by the present situation, so your hungry pet speaks like, well, a hungry pet. You may even set it to describe a notice aloud so you don’t have to glance down.

Firmware is built with ESP-IDF, and on the phone side, we utilize Kotlin with BLE, Opus, Whisper.cpp, Llama.cpp, Piper, and espeak-ng for phonemes, all because we need them. The source code is released under the Apache 2.0 license. Unfortunately, because of the phoneme connection, the released APK becomes GPL-3.0, thus if you ship tagged binaries, you must also supply the matching source. To build, you must flash the board, assemble the program using JDK 21, retrieve all native source, drop prebuilt libraries into place, and keep your firmware and app on the same commit to ensure that the wire protocol matches. If you receive a black screen after a flash, a simple power cycle will generally fix it.
[Source]