All projects

ORKA

An Android task app whose reminders are full-screen alarms with Start, Done and Dismiss, backed by a language model running on the phone.

Kind
Android app with an on-device model
Year
2026
Built with
  • Kotlin
  • Room
  • AlarmManager
  • LiteRT-LM
  • Gemma 4 E4B

The question

What makes a reminder hard to ignore without making it annoying?

ORKA puts the alarm on the screen with the actions you need, and learns from how you answer it.

Why it exists

A reminder is easy to dismiss. ORKA schedules exact alarms that take over the screen, offers actions to start, finish, reschedule, snooze or split the task, and records how you respond.

How it works

From a sentence to an alarm that adapts

  1. 1

    Task

    You type something like “call the bank tomorrow at 4”.

  2. 2

    Draft

    The on-device model turns it into a draft with a time. A set of rules does the same if the model is not available.

  3. 3

    Alarm

    You confirm, and an exact alarm is scheduled that can take over the screen.

  4. 4

    Response

    You tap Start, Done, snooze, reschedule or split, and each choice is recorded.

  5. 5

    Learning

    The scheduler uses those records to place the next reminders.

Decisions

  • The model stays out of the app package

    The model file is pushed to the phone once. Builds stay small, the model survives a reinstall and it is picked up without a restart.

  • Every model call has a rules fallback

    Parsing, splitting, rescheduling, snoozing and clarifying each fall back to rules if the model is missing, too small, too slow or returns something unreadable.

  • The learning scheduler waits for enough data

    Reinforcement learning takes over only after 40 interactions and 10 completions. Until then the rule-based and adaptive modes schedule.

  • Backup keeps your tasks but not the model

    The model copy is excluded from backups because it is large and can be pushed again, while the task database is included.

Evidence

  • Built

    About 87 test methods, 66 that run on the JVM and 21 that run on a device, with database schemas exported for versions 1 to 4 and a migration test.

Built with

Languages

  • Kotlin

Frameworks and libraries

  • Room

AI and models

  • LiteRT-LM
  • Gemma 4 E4B

Systems and hardware

  • AlarmManager