PERSONAL ASSISTANT · 2026
Jana
An assistant that lives in Telegram under an account of her own and answers like a person, not like a bot with slash commands.
beta launching September 2026
Why
For several months I watched the assistants my friends were running: I tried talking to them, and listened to what people complained about and what they praised. The conclusion came out the same every time — setting one up is hard work, and half the built-in features are of no use to the owner. So I decided to build my own. I wanted to build my own memory framework and my own functions, and to get right into it.
How I did it
The model itself remembers nothing, sees nothing and can do nothing with its hands. All of that comes from the scaffolding around it, and the scaffolding I built. The first things working were voice transcription, memory and a morning brief with the day's tasks; right behind them came web search, her own browser, reading images, and a calendar with reminders. Jana lives under a full Telegram account, which is why she joins chats and groups and subscribes to channels on her own — a bot cannot do that.
What she can do
-
Watches production and answers clients
She has access to my services in production: she reads logs live and tells me herself if something has fallen over or a client has hit an error. She is also the first line of support — clients write to her directly, she answers herself, and calls me in only when she does not know the answer.
-
Hears and sees
Voice messages are transcribed on my own laptop through mlx-whisper — free, and the recording never leaves the machine. A photo she is sent is shown to the model and deleted immediately. She sees stickers and the reactions under messages.
-
Remembers
A separate context per person, carried on between messages. She remembers who I agreed to meet, what I was thinking about last week, and what my ideas turned into. Plans for posts and projects I simply dictate to her out loud.
-
Works with her hands
A voice message that starts with the word «task» switches her into working mode: she edits code and fixes projects straight from the chat while I am out walking. First she restates the task in her own words, so I can see whether she understood me before she touches any code.
-
She does not lie about what she has done
When Jana reports on a job, the report is checked against reality: what appeared on disk, what changed in the project, which tools she called — and any discrepancy is written into the answer. If a reply says «noted» or «I will remind you» while no reminder actually exists, the reply simply does not go out: she redoes it. A model will happily say «done», so a task is closed by the fact, not the phrase.
-
Comes to me first
In the morning she sends a summary and the day's tasks. In the evening she asks whether I did them, and carries whatever is left over to tomorrow. She reminds me about training, about meetings, and about what needs preparing for a meeting.
-
Assembled from modules that match a life
Separate modules for calendar, tasks, food diary, health, sport, psychology and photo search. Each one can be edited or switched off on its own — which stops working once everything lives in one sprawling prompt.
-
Keeps a medical record
Appointments, test results, how indicators move over time, prescriptions — all in one place and linked by date. Of all her functions this is the one people who have seen Jana working most often ask to have for themselves.
-
Obeys only me, and restarts herself after a crash
An allow-list by identifier, a per-person hourly limit and a daily one. In group chats she answers only when called. «This is Daria» from somebody else's account does not get through. If the process dies, a watchdog brings it back without me.