ComingUp ComingUp
RoutineArc — on-device AI planner for Android

RoutineArc — on-device AI planner for Android

Android day planner with on-device AI — no backend, no per-user AI costs.

jasonfras90 Posted by jasonfras90

Is this your project? Ask us to take it down

Aug 18, 2026 Other

Gallery

RoutineArc — on-device AI planner for Android

About

RoutineArc is a working Android productivity app built around one idea: the AI

should run on your phone, not on someone else's server.

It is not a finished product. It is a substantial, tested codebase that runs on

real hardware and still has a known list of work to do before it could ship —

that list is at the bottom of this listing, not hidden.

You capture the way you think — type "dentist tomorrow at 3pm", say it out loud,

or photograph an appointment letter — and it becomes a properly dated task or

event with priority, duration, location and repeat already filled in. It then

plans your day around the calendar events you already have.

THE ENGINEERING THAT MAKES IT WORTH BUYING

Three interchangeable AI backends sit behind a single interface, and the user

picks which one runs. All three are scored against the same 27-case, 91-field

capture fixture:

• Gemini Nano — uses Android's built-in AICore, nothing to download.

91/91 at best; 88–89/91 under sustained use, because AICore rate-limits

repeated calls and the run retries.

• Gemma 4 E2B — one-time 3.68 GB download, runs on the Tensor G5 NPU.

90/91, verified on device. Recorded at 91/91 earlier; re-measured across four

runs on 15 Aug 2026 it scored 90/91 every time, always losing the same field.

• Rules engine — needs nothing at all.

91/91, verified in the test suite. Deterministic, so it is reproducible in CI

rather than needing a phone; individual captures were also confirmed on device.

That third one is the commercially important one. The app still works on phones

that can run neither model, with no cloud fallback — so there is no inference

cost per user, ever, and no server to keep alive.

Because there is no backend, there is no hosting bill and no scaling work.

Everything lives in SQLite on the device. The app makes no network calls of its

own except an optional address lookup that ships switched off; choosing Gemma 4

downloads the model once, and tapping "Navigate" hands an address to Google Maps.

WHAT'S IN IT

Capture by typing, voice, or photo with on-device OCR · AI day planning around

real events · Eisenhower matrix with drag-to-reprioritise · focus sessions

(Pomodoro, timer, stopwatch) · habits with streaks · weekly review, evening

triage and a "Keep It Honest" screen · saved places with "leave by" travel times

· work schedules and shifts · recurring tasks · home-screen widgets · Wear OS

companion · PDF export · backup, restore and CSV import · a built-in user guide

that exports itself as a PDF.

28 screens.

MEASURED, NOT ASSERTED

Every figure below came off a real device or a real build, and the method is

recorded in the engineering audit inside the repo:

• Capture accuracy — the three scores above, against a shared fixture.

• Cold start 251–287 ms on a Pixel 10 Pro XL, force-stopped between runs.

• Scroll jank 0.00% on the calendar at 30 tasks and 8 events; no missed vsync

on any screen measured.

• Download size about 61 MB per device on arm64, built as an App Bundle.

• 2,141 tests across 167 suites, running in about 15 seconds from a clean clone.

Where an earlier claim turned out to be wrong, it was corrected rather than

quietly dropped — the audit records the withdrawn ones too.

BUILT TO BE HANDED OVER

• 2,141 tests across 167 suites — including the data layer tested against a real

SQLite engine, and a smoke test that mounts every one of the 28 screens.

• A custom Kotlin native module exposing Gemma 4 on the Tensor NPU through

LiteRT-LM. This was the hard part.

• A measurement harness that scores capture accuracy against a fixture, so any

future change to the AI can be proven not to have regressed it.

• A guided code tour written for whoever owns it next: where to start, why the

unusual parts are the way they are, and the traps that would otherwise cost a

day each.

• An engineering audit recording what was measured, what was fixed, and what was

deliberately left alone.

TECH STACK

Expo SDK 54 · React Native 0.81 · TypeScript 5.9 · expo-router · SQLite

(expo-sqlite) · ML Kit GenAI (Gemini Nano) · ML Kit text recognition (OCR) ·

LiteRT-LM + custom Kotlin Expo module (Gemma 4 on NPU) · Jest

TRAFFIC AND REVENUE

None. Never published to Google Play. No users, no downloads, no revenue. This

is a working codebase with known remaining work, not a running business.

WHAT'S INCLUDED

Full source repository (private GitLab, access granted on agreement) · the custom

Gemma NPU native module · 2,141 tests · capture measurement fixtures and harness

· in-app user guide and its PDF generator · guided code tour, engineering audit

and release documentation · app icons and branding · copyright assignment on

completion.

KNOWN ISSUES AND LIMITATIONS — DISCLOSED UP FRONT

• Release builds are currently signed with the debug keystore, so a real signing

key must be configured before Play submission. Documented; roughly an hour.

This is the only hard blocker to publishing.

• Tested on one phone model. All device verification was done on a Pixel 10 Pro

XL, plus one Android emulator used to confirm the app degrades correctly where

no AI hardware exists. It has not been run across a range of screen sizes, OEM

skins or Android versions. minSdk is 26.

• The Android build directory is not in the repo. A fresh clone runs

`expo prebuild`, then re-applies two documented manual edits: the

uses-native-library declarations that give Gemma 4 access to the NPU, and a

signing key. The README covers both. Budget an hour before your first release

build.

• The universal APK is 305 MB because it bundles four CPU architectures.

Building as an App Bundle instead delivers about 61 MB per device on arm64,

with no code change — measured from the built bundle, not estimated.

• Photo capture scores 31/32 on the image fixture where text capture scores

90/91. One field remains: on a patient-portal screenshot the model names the

clinic in the notes rather than the location field. The other five fixtures

pass completely.

• Two written and tested modules are not wired up: share-sheet intake, and

location-based reminders. Both need native edits in the generated Android

directory; the audit explains why that was deferred rather than risked.

• The Wear OS companion sends data but has not been verified against a physical

watch.

• Development was AI-assisted. A large amount of this code was written with an AI

coding assistant, under my direction and review. I am saying so up front

because you would reasonably want to know. The answer to the obvious next

question — whether it is held together — is the part I would rather be judged

on: 2,141 tests, a data layer exercised against a real SQLite engine, a smoke

test that mounts all 28 screens, and AI accuracy measured against a fixture on

real hardware rather than asserted. The engineering audit in the repo lists

what is still open, including the items above.

A note on the screenshots: a few carry small grey rectangles. Those are

deliberate redactions covering my own real tasks and appointments — the shots

were taken on a device in daily use. They are not UI defects.

Comments (2)

Kenya Huels Kenya Huels 3 months ago

zero per-user ai costs is smart, which model runtime

Theodore Herman Theodore Herman 2 months ago

running inference locally is the right call for privacy