How Tasuke AI turns one sentence into several tasks
Most to-do apps want a form: a title field, a date picker, a time picker, a save button — once per task. That is fine at a desk. It is useless when you are walking, driving, or halfway through something else and three things occur to you at once.
Tasuke AI takes the opposite approach. You say everything in one go, the way you would say it to a person:
"Tomorrow at 3 PM send the build to Emily and Friday check App Store"
and you get this back to confirm:
| Task | Date | Time |
|---|---|---|
| Send the build to Emily | tomorrow | 3:00 PM |
| Check App Store | the coming Friday | — |
Here is what happens in between. All of it runs on your phone.
1. Listening, without a recording
While the recording screen is open, audio from the microphone is streamed straight into a speech-recognition model that ships inside the app. It is never saved as a file, and it is discarded as it is processed. When you stop, what remains is text.
The model is an English speech model running on your phone's own processor, so there is no upload step and no server waiting on the other end. That is also why the app works in airplane mode.
2. Finding where each task ends
Real speech is messy. People say "um", change their mind mid-sentence ("at 4 — no wait, at 5"), string tasks together with "and", "then" or "also", and sometimes use "and" inside a single task ("ask Bobur and Aziz to confirm the dinner").
Tasuke AI splits the transcript into clauses with a set of language rules, then decides which clauses are separate tasks and which belong together. Filler words are dropped, and each title is cleaned up so it reads like something you would have typed: "Send the build to Emily", not "and um send the build to Emily".
3. Reading the dates and times
Each task is then checked for when: "tomorrow at 3", "Friday", "next Monday", "in two weeks", "the 14th", "tonight". Relative dates are resolved against the moment you spoke, so "Friday" said on a Wednesday means the Friday two days later. A date said once can apply to everything in the sentence — in "tomorrow, pack the suitcase and exchange some dollars", both tasks are for tomorrow.
Some conventions are fixed so the result is predictable: "tonight" means 8:00 PM, and a task with a date but no time is an all-day task.
4. You check it
Speech recognition can mishear a word, and any automatic reading of language can pick the wrong Friday. So nothing is saved until you have seen it. The confirm screen shows every task with its date and time, and you can fix or remove anything before you tap Save.
5. Reminders, scheduled by your phone
Once saved, tasks go into a database inside the app's private storage, and every task with a time gets a local notification scheduled by your phone's operating system. There is no push server involved, which is why a reminder still arrives when the app is closed.
Why rules, and not a chatbot?
The speech step uses a neural model. The splitting and date reading do not: they are deterministic rules.
We tried a compact on-device language model for the extraction step. On notes it had never been tuned on, the rules produced the exactly right result far more often — 75% against 32% in our test at the time — and the model was removed. Rules have two other properties we care about: the same sentence always gives the same result, and they cannot invent a word you did not say.
The rules are held to a large set of example notes in the app's test suite, and a change that fixes one phrasing but breaks others fails the build. The sentences in the demo on our home page come from that suite, and each result shown is what the extractor produces for it.
What it does not do (yet)
It understands English only, because the built-in speech model is an English model. And it is not perfect — long, rambling notes with several dates are still the hardest case, which is exactly why the confirm screen exists.