Where the idea came from
I live in Belgium. The town hall, the doctor, the landlord, the bank: almost every important conversation happens in Dutch or French, not in my language. A phone translator helps with a sign or a letter, but in a live conversation it gets in the way. You tap buttons, pass the phone back and forth, and hope the numbers and dates came out right.
On 17 October Tolkimogi began as an experiment for the Telegram channel «Бельгия на связи» (Belgium on the line). By the 18th I saw it was a product of its own and gave it a repository. By 29 October it holds 544 commits. As with my other projects, an AI agent writes the code; I set the tasks, check the results and make the decisions.
What Tolkimogi does
You put the phone on the table between two people. Each speaks their own language, and the interpreter translates aloud both ways. It is not dictation; it is a voice that keeps the conversation going the way a human interpreter would.
it asks again when it did not catch a date, an amount or a number;
it can quietly explain to you what was meant without interrupting the other person;
it speaks on your behalf when you ask it to say something;
it softens rude words instead of translating them literally.
What a conversation looks like
Before you start, you choose who you are talking to and in which language. After that you just talk. The app has two interpreter voices, Eva and Max, and the project mascot is Felix the cat. He also stars in the voiced cartoon on the site: 54 seconds about a town hall visit with an interpreter.
Under the hood is OpenAI's GPT-Live voice model: it hears speech and answers in voice right away. The mobile app is built with Expo and React Native, and a Fastify server runs in Docker on my VPS. iOS builds go through EAS to TestFlight; I build Android locally.
What broke
A lot broke in 12 days. Here are the four cases that taught me the most.
21 October. An experimental audio path the agent added for quality cut off the microphone: the model could not hear me, and the call dropped after about 15 seconds. I switched the new path off and went back to the old one.
The same evening the interpreter said “I'll translate everything, one second” and went silent. An overly cautious prompt told it to “check the language with the client”, so it waited instead of translating. I was charged for 2 minutes. A new prompt of six short rules got the same scene right, including “950 € a month” and “75 € bijkosten” (service charges).
On 22 October I tested a cheaper chain: speech recognition, then a text model, then speech synthesis. It was just as accurate, but answered in 3.8 seconds instead of 1.7, and in noise it merged lines from different people. I decided not to pursue it.
On 28 October a tester's network dropped right at the start, and their account was stuck for 2 hours on the “This meeting can't continue right now” screen. We fixed it the same evening.
The main lesson: for live interpreting, response speed matters more than saving money. One second of silence is fine; four is not.
Privacy and history
A conversation with a doctor or a landlord is personal. So Tolkimogi has no ads, and conversation history is stored encrypted and deleted after 90 days by default. You can open it to reread what you agreed on.
Where the project stands
iOS: an external TestFlight group, build 1.0.0 (9) in testing. As of 27 October the group has one tester and 6 sessions; the group limit is 30 people.
Android: internal testing on Google Play. To go public from a personal developer account, Google requires a closed test with at least 12 testers for 14 days in a row.
No App Store sales yet: in-app purchases are not set up.
tolkimogi.com is live: five languages, a test waitlist and 30 free minutes.
Money
There is no revenue, and you cannot pay yet. A minute of conversation costs me about 5 cents at OpenAI, so the free test is capped at 30 minutes. After the test you will pay per minute, with packs or a monthly plan. I will announce prices before launch; first I want to see how many minutes a real conversation takes.
What's next and how to try it
The main task now is real testers: people who actually have a conversation ahead of them in a language they don't speak, at the town hall, the doctor's or with a landlord. Their feedback will show where the interpreter makes mistakes and what one real conversation costs.
You can join the free test at tolkimogi.com; the link is on the project page above this article. The next report will cover what testers found in their first conversations.