Joins your meeting, transcribes and translates the conversation. Say "Kâtip" and it takes a note, records the decision and answers your question from the meeting's own record.
Kâtip is a voice AI assistant that joins online meetings. It transcribes the conversation live, translates it into another language and tells speakers apart. When addressed by voice or in the meeting chat, it records notes, decisions and action items and answers questions in writing or by voice. When the meeting ends it prepares a Word and PDF report.

A decision made in a meeting often lasts only as long as the attention of whoever takes notes. Who took on what, by which date, and why the decision was made either stays in someone's notebook or is never written down. Even when a recording exists, nobody listens to hours of audio afterwards.
Automatic note takers close part of the gap but lose trust in two places. The first is who spoke: a tool that guesses when tying a voice label to a name can attribute words to the wrong person. The second is answers: a tool that answers a question about the meeting with something that is not in the record does more harm than taking no notes at all.
Teams working in Turkish face one more problem. Features of the language itself, such as forms of address, suffixes and how numbers are read aloud, are an afterthought in most tools.
Kâtip joins the meeting as a guest, under the name the organisation sets in the panel. It is sent from the admin panel with the meeting link, the participant list and optional documents, and waits in the lobby until a participant admits it. The meeting audio is collected into a single stream; transcription, translation and speaker separation happen live on that stream.
Addressing it with something like "Kâtip, record this as a decision" puts a note, decision or action item (with owner and due date) on record; the same requests can also be typed into the meeting chat. It answers questions from the meeting transcript, the notes taken and uploaded documents (PDF, PowerPoint, Word, text). Asked about something that is not on record, it says "I have no record of that"; answers drawn from a document come with the file and page as the source.
By default it writes its answer into the meeting chat, and it can answer by voice on request. When speaking it does not interrupt: it waits for the person asking to finish, stops if someone starts talking, and writes the answer to the chat if the audio channel is busy.
When the meeting ends a Word and PDF report is prepared: executive summary, agenda, decisions, action items (owner, due date, priority), open questions, notes taken by Kâtip, participant contributions and screenshots. A signed, time limited link to the report is posted in the meeting chat.
The voice labels speech recognition produces are anonymous and never count as an identity on their own. Writing a name against a sentence requires two conditions together: a single person must appear as speaking on the platform at that moment, and that person must have been explicitly bound to an email address in the panel. If either is missing, the sentence stays as "Speaker 2"; Kâtip does not guess.
The language model can add notes, decisions and action items and can write to the chat. Deleting a record, pausing recording, taking a screenshot, changing the translation language and leaving the meeting run only on an explicit command and are never left to the model's judgement. Anyone can suggest a standing instruction for Kâtip, but it takes effect only when a company account signed in to the panel approves it.
Kâtip answers only from that meeting's record, the notes, the participant list and the uploaded documents. An answer from a document that cites a source that was not retrieved is rejected. The transcript and documents are given to the model as untrusted data; an instruction embedded in them is not treated as a command.
Addressing recognises Turkish inflections ("Kâtip", "Kâtibim", "Kâtip'e") and tells talking about Kâtip apart from talking to it: "Kâtip de not alıyor" (Kâtip is taking notes too) is not a command. Amounts, dates, times and percentages are spoken with their Turkish reading in voice answers. The target language of live translation can be changed by command during the meeting.
When Kâtip joins it posts the KVKK (Turkish data protection law) notice to the chat and sends a short reminder to late joiners. In "off the record" mode the text of that part is not stored and nothing is sent to the language model. No audio recording is kept. When a person is deleted in the panel, their name is anonymised in past transcripts and the related report files are deleted. Speech recognition and the language model run at external providers; because audio and text are transferred to them, we recommend that the data transfer is reviewed with the organisation's legal team before installation.
Kâtip does not use an off the shelf meeting bot service; it has its own browser based joining layer for Teams, Meet and Zoom. Speech recognition does not start while the bot waits in the lobby. Limits set in the panel cap model requests per meeting, concurrent bots and total daily time, and usage and cost estimates are tracked per meeting.
Kâtip is at version 0.1.0 and in pilot use. It has been tested live in real Microsoft Teams meetings: transcription, voice answers, not interrupting, note taking, leaving the meeting and report generation. Google Meet and Zoom connections are written; their live validation is next.
The admin panel brings together sending Kâtip to a meeting, system health, past meetings and reports, usage and cost estimates, settings, people and security events. Panel sign in is role based (administrator, operator, viewer), with a brute force lockout, CSRF protection and strict security headers. The product is tested by more than three hundred and fifty automated tests.
On the roadmap: email integration with Microsoft 365, enterprise single sign on (SSO), live validation of Meet and Zoom, and bringing to production the setup where every meeting runs in its own isolated worker.
If you would like to try Kâtip in your own meetings, fill in the form. Let us talk about which meeting platform you use, your team's working language and your data retention expectations.
Talk about a pilot