A private AI meeting assistant records, transcribes and summarizes meetings without handing your audio to a cloud product you don't control. This is the story of RecordAI, the local-first meeting assistant I designed and built for my own product-discovery work: why the default tools failed, what privacy actually means in the architecture, and why Hebrew had to be day-one, not a feature request.
Every product-discovery call starts with trust. A founder sits across from you and describes the thing they haven't launched yet: the idea, the numbers, the doubts. It's some of the most sensitive material a small company has.
And the standard way to take notes on that conversation is to invite a third party into it. A cloud meeting bot joins the call, the audio goes to servers you don't control, under terms you didn't read, and the founder's unlaunched product becomes training-adjacent data somewhere. Everyone has learned to say "hope you don't mind the bot". I minded.
The privacy question isn't paranoia, it's professional courtesy. If I ask people to trust me with their product at its most fragile stage, the least I can do is know exactly where the recording of that conversation lives.
I tried to make the existing tools work before building anything. They were weakest exactly where my work needed them most.
Hebrew was the first wall. My discovery calls are in Hebrew and English, often inside the same sentence. The big tools are English-first with Hebrew as a language pack, and a Hebrew transcript that almost works is worse than none: you stop trusting it, so you stop reading it, so you're back to scribbling.
Then the shape of real meetings. Some of my conversations happen in a room, not on Zoom - and cloud bots only know how to join calls. And when summaries did arrive, they were generic minutes. After a discovery call I don't need "the participants discussed the roadmap". I need insights, decisions, open questions and next steps, in the shape I'll actually use an hour later.
So I built RecordAI around one rule: the original audio never leaves my machine.
In a room it behaves like a local recorder. On a call it captures my microphone and the system audio separately, which gives an honest split between "me" and "everyone else" - a deliberate choice, because I'd rather have a separation that is always true than a promise of perfect speaker recognition that sometimes invents things.
Only one step leaves the machine: transcription, and it runs on a private worker I set up myself, built with IVRIT/faster-whisper. Long recordings are chunked, transcribed and merged back into one continuous timeline, so the plumbing never becomes the user's problem. Hebrew and English are in the core architecture from day one - not an add-on, not a beta flag.
Keeping the source audio local does something else too: the trust boundary is visible. When a summary claims a decision was made, the original recording is right there to check against. The AI can be wrong. The tape doesn't lie.
Recording and transcription are plumbing. The part that decides whether a tool survives in a workflow is what you get afterwards.
RecordAI's summary is designed around the decisions a product designer makes after a call: what did we learn, what did we decide, what's still open, what happens next. The AI output is part of the interface, not a block of text you have to re-read and rewrite into something useful.
Preparation got the same treatment. The point of a discovery call is what you want to learn, so the tool starts before the meeting: walking in knowing the questions. Recording itself is deliberately boring - calm and obvious - because the moment a recorder needs attention, it's stealing it from the conversation.
RecordAI fits how I meet: Hebrew and English, rooms and calls, sensitive conversations, follow-up in the shape I work in. But the reason I'd tell this story to a founder is the general lesson.
A useful AI product isn't defined by having AI. It's defined by privacy, cost, language quality, reliability and control over the workflow. The model is a component; the product is everything around it - where the data lives, what the output is shaped for, and whether the boring parts are boring enough to disappear.
That's also how I approach products for clients. If you're building something with AI inside, the model choice is rarely your real problem. The trust boundary and the workflow are. The full build story, with screens and decisions, is in the RecordAI case study.
A private AI meeting assistant records, transcribes and summarizes meetings without sending your audio into someone else's cloud product. In RecordAI's case, the original audio never leaves the local machine; only the transcription runs remotely, on a private worker set up for that single job.
Most meeting tools are built English-first, with Hebrew added as a secondary language pack. The result is transcripts that look usable until you try to quote them. RecordAI's transcription runs on IVRIT/faster-whisper, so Hebrew and English are part of the core architecture from day one, not an add-on.
Most cloud bots only join video calls. RecordAI records in-person meetings through the microphone and keeps the original audio locally, so a conversation in a room gets the same preparation, transcription and follow-up as a remote call.
The original audio stays stored locally on the machine that recorded it. That keeps the trust boundary visible, and the source is always there when a transcript or a summary needs checking against what was actually said.
Generic minutes log what was said. A discovery summary is shaped around what a product designer does after the call: insights, decisions, open questions and next steps. The AI output is part of the interface, not a block of text you re-read and rewrite.
On calls it captures the microphone and the system audio separately, which honestly separates 'me' from 'everyone else'. That is a deliberate design choice: no claim of perfect person-by-person speaker recognition, just a split that is always true and always useful.
RecordAI was designed and built by Itamar Katan, a freelance EdTech product designer and AI builder, for his own product-discovery work. The existing tools were weakest exactly where the work needed them most: usable Hebrew, reliable local recording, meetings that happen in a room, and summaries built for product discovery instead of generic minutes.