Bloobert

Problem

People sit at the computer for hours without breaks. I built Bloobert, a translucent jelly pet that melts while you sit and regains its shape when you take an active break. I defined success as using it myself every working day. I did no user research.

Constraints

• Windows 11 first, macOS next.
• Local only: no account, no telemetry. Idle detection reads only a timestamp, never which keys were pressed.
• A small, transparent, always-on-top window (252×182).
• Unknown presence must not advance the sitting clock: melting unfairly is worse than melting late.

Decisions

• Tauri 2 over Electron. It reuses the existing SVG and WebAudio prototype, Rust reads the idle time, and the footprint is small. A spike found no failure criterion that required Electron. I didn’t measure an Electron baseline.
• Procedural sound. WebAudio in TypeScript: one triangle oscillator per vowel, a pitch drop, a lowpass filter and random jitter, lower and slower as Bloobert melts. It needs no voice actor, carries no voice licensing and works in any language.
• SVG rendering. A sticker look ported from a design study. Three.js 3D is the plan B.
• Layered break detection. Raw idle time cannot tell whether you are seated. The layers are an input grace period (5 minutes proposed), then whether media is playing, then an opt-in camera during breaks only.
• Safe saves. Writes are atomic. An unreadable or unknown-version save file is renamed, never overwritten.

What I owned

Everything, working with coding agents: the product definition, the architecture, the Rust and TypeScript code, the design and the tests.

Result

• It runs as a desktop app. The spike measured a working set of about 37 MB, in a debug build, “observed roughly”. I haven’t measured release size or CPU.
• Rust tests grew to 126 and the frontend suite passes 11 of 11.
• Five deliberate mutations of the break rule each made its tests fail.
• There is no release or installer yet.

What I’d change

• I’d measure a release build (size, CPU), and check the Windows transparent-window border and click-through behavior, which the docs list as still to be checked.
• Without the camera, break detection is trust-based. I still need to decide whether that is acceptable or needs another signal.