LaundryLens

From a pile of laundry to wash rounds

Sorting a laundry pile by hand means re-reading labels, grouping similar fabrics, and deciding what can safely share a wash. LaundryLens starts from a photo instead: you photograph the pile, and the app lists the items it finds — a name, a fabric family, and a colour group for each — with a photo crop of every detected item so you can check the detection against the original picture.

Detected items stay editable. Each item can be corrected by hand, including its care-label reading through a dedicated dialog, and earlier scans remain in History. A deterministic Kotlin rule engine then builds up to four wash rounds from the reviewed items, choosing the order of rounds and each round's temperature, spin, and duration, with visible reasons for every recommendation.

Seeing on the phone, deciding in Kotlin

The work is split deliberately. Perception is the model's job: a Gemma 4 E2B vision model runs entirely on the phone through LiteRT-LM and lists the items in the photo. The wash plan is not the model's job: a deterministic Kotlin rule engine turns the reviewed items into rounds, so the same item data produces the same plan and every recommendation carries an inspectable reason instead of a language-model opinion.

The model file is a one-time, in-app download of about 2.6 GB, transferred over HTTPS and SHA-256-verified before it is used. After that, scans work offline. The app also works without the model at all: in manual mode you enter items yourself and the same rule engine plans the washes.

Smart Mode is the one optional network feature and it is off by default. When you enable it, the photo you scan is sent to my own server for item classification, and any failed request falls back silently to the on-device scan. There is no account, advertising, or analytics, and in-app telemetry stays on the device.

Early beta and its limits

This is an early beta. The English-only interface runs on Android 8.0 and later, and the published download is the signed v0.3.1 (code 11) phone APK with its exact byte size and SHA-256 shown below. Screens on this page are emulator captures of demo scans on staged laundry photos; they show the interaction, not measured recognition accuracy or care outcomes.

• Scan time: on the emulator test builds, CPU-only scans of a pile took roughly 20–45 seconds; GPU execution is expected to be faster but is not part of this release.
• Recognition quality: the demo scans are not a benchmark. Detection quality on physical devices, camera angles, and real wardrobes is not established by this page.
• Source: the app is MIT-licensed, but the corresponding source is not published yet, so this release shows no source link; byte-level verification runs through the published checksum.

The release checks were emulator-based: upgrade installs, scan lanes on staged photos, Smart Mode on and off with endpoint-failure fallback, and manual mode. They establish that the shipped flows execute, not long-term physical-device reliability.