Thirty seconds back, thirty seconds forward.
A ping without context is a jump scare. Every MeetPing alert opens a popover with the last 30 seconds of confirmed transcript before the keyword fired, and starts collecting the next 30 seconds live. You read backward, then forward, then answer.
Meetings move faster than they read. By the time someone says your name, the question that prompted it is twelve seconds old and you are six sentences behind. The traditional fix is "sorry, could you repeat that?" — which costs you trust and costs the meeting a minute. The MeetPing fix is to keep a rolling 30 seconds of audio and confirmed transcript in RAM, snapshot it on the keyword hit, and show it to you alongside the next 30 seconds as it streams in.
Why 30 + 30 specifically? It maps to working memory. People keep about half a minute of conversational thread in their head; below that you're rewinding the meeting, above it you're reading rather than catching up. Six hundred words of essay on this calibration is over here.
okay let's pull up the metrics from last sprint and walk through them
the activation chart is up 14% week over week, retention basically flat
we did get a bunch of churn from the enterprise tier, mostly billing-related
so on the next quarter — Ogtay, what does the
roadmap look like for the listener service?
do we still want to ship the remote alerts channel before the holidays?
and how are we thinking about the privacy posture of on-device Parakeet
right, makes sense — pin that for the security review thread
How the buffer actually works
The audio side is a 30-second ring buffer of 16-bit PCM frames living in the MeetPing process. The text side is a parallel ring of confirmed Parakeet tokens with timestamps. When a keyword fires, both rings are snapshotted; the audio side is released back to the listener (we keep the ring rolling for the next ping) and the text snapshot becomes the popover's "past" column. The "future" column streams in as Parakeet confirms tokens after T+0; it stops appending after 30 seconds.
Two corollaries that matter. One: the past column is fixed at the moment of the ping, so reading it is reading what was actually said up to that point — not a re-run of the model on re-played audio. Two: nothing about this touches the disk. The buffer ages out as new audio comes in, and the snapshot only lives as long as the popover is open.
What you do with the next 30 seconds
The future window is the difference between MeetPing and a note-taking tool. A note tool tells you what was said. MeetPing shows you what's being said while the speaker is still saying it. You can read the question forming, answer when it actually lands, and look like you were tracking the whole time. We refer to this as "the receipt." It's the receipt that justifies $24.90 lifetime.
Keyword watch
A list of names and topics, matched live against the Parakeet stream. Pings the moment one fires.
Profiles
Per-context watchword presets. Engineering, sales, all-hands, parental. One click to switch.
On-device privacy
Zero outbound calls. RAM-only audio. No transcript file. Pulled from a real audit.
Catch up without rewinding.
Every ping ships with thirty seconds of context behind it. The next thirty fill in live, in the popover.