We built Craqly: real-time AI help during the call, not after it

#ai #career #productivity #showdev

Every meeting tool we tried solved the wrong moment. You finish a sales call, and twenty minutes later a beautiful summary lands in your inbox telling you exactly which objection you fumbled. Thanks. The call is over.

So we built Craqly to work during the conversation instead.

What it does

Craqly is a desktop app for Mac and Windows that listens to whatever call is on your screen (Zoom, Google Meet, Teams, WebEx, Slack huddles, even a podcast in a browser tab) and surfaces short, context-aware suggestions in real time. Latency is 1-2 seconds for text answers.

Three things make it different from the recorder-bots:

No bot joins the call. It reads system audio directly, so there is no awkward "Craqly Notetaker has joined the meeting" moment and no participant for procurement to object to.
The overlay is invisible to screen-share. The window is excluded from screen capture at the OS level (CGWindowSharingNone on macOS, WDA_EXCLUDEFROMCAPTURE on Windows), so when you share your screen the other side sees nothing.
It answers during the call. Sales mode spots objections and pulls the counter from your battle cards. Interview mode anchors answers in your resume. Meeting mode captures action items with owner and deadline while they are being said.

The hard parts

The honest engineering story: real-time is unforgiving. A summary bot can take 30 seconds; nobody notices. We had to get from audio to transcript to LLM response fast enough to be useful mid-sentence, which meant streaming STT with interim results, aggressive prompt assembly, and one nasty echo-dedup problem when the mic picks up the speaker output and every line shows up twice attributed to two different people.

Transcription runs in 18 languages with mid-conversation switching, which sounds like a checkbox feature until a Hyderabad standup flips between English and Telugu four times in one sentence.