How It Works

The mechanics of the Google Meet integration — transcription, voice triggers, TTS responses, and post-meeting summaries.

This page is the technical companion to Google Meet Overview. It explains how Cygent actually participates in a call — what it listens for, when it speaks, and what happens after.

The lifecycle of a Meet session

Real-time transcription

Cygent transcribes the full call continuously. This isn't just a recording — it's a live stream of text that feeds Cygent's reasoning in real time. When you ask a voice question, Cygent already has the full conversation context up to that moment. You don't need to re-explain what you were discussing.

The transcription is what powers the post-meeting summary too — Cygent isn't reconstructing the meeting from memory, it has the full verbatim text.

Voice activity detection and name triggering

Cygent doesn't respond to every utterance. It listens specifically for its own name as the start of an addressed phrase. Examples that will trigger:

• "Cygent, what do you think about this design?"
• "Hey Cygent — can you explain finding M-3?"
• "Cygent, is that exploitable given our protocol constraints?"

Examples that will not trigger:

• "I asked Cygent yesterday and..." (name appears mid-sentence, not as an address)
• "We should have Cygent review this later" (referring to, not addressing)

This heuristic keeps Cygent from interrupting. The cost is the occasional missed question when someone phrases an address unusually — in which case just rephrase.

TTS responses

When Cygent replies over voice, it uses text-to-speech. Responses are kept short on purpose — typically 20 to 60 seconds — because no one wants a bot monologuing on their call.

For anything longer, Cygent says something like "I've dropped the details in the chat" and posts the long-form content into the Meet chat instead. The voice stays tight; the chat carries the depth.

Supplementary chat

Some things don't read well out loud — file paths, code snippets, line numbers, URLs. When Cygent has those, it posts them in the Google Meet chat in parallel with its voice response. So a voice reply like "I'd take a closer look at getCollateralValue in Vault.sol" gets paired with a chat message showing the exact line range and a link back to the finding detail page on the dashboard.

Post-meeting summary

When the meeting ends, Cygent produces a structured summary:

• Key topics discussed. Major themes, with 1–2 lines each.
• Decisions made. Concrete outcomes, including any accepted risks.
• Action items. Tasks identified during the call, with the assigned participant's name where known.
• Follow-ups. Anything Cygent should do next — file an issue, run an audit on a specific path, create a scheduled task.

The summary lands in the channel or thread that triggered the Meet join. If you want it to land in a specific thread, invite Cygent from that thread.

Per-participant DMs

Alongside the channel summary, each participant gets a direct message from Cygent with:

• Their personal action items from the meeting.
• Context or follow-ups relevant specifically to them (e.g. "you mentioned you'd take the oracle refactor — here's the relevant finding").

This works because of speaker diarization during transcription — Cygent knows which participant said which thing and can route personal follow-ups accordingly.

Context capture for future audits

Anything said in the call that matters for the protocol's security posture gets lifted into Cygent's persistent memory. A decision like "we're going with Aave V3's liquidation math" becomes a rule that future audits check against. An accepted risk like "we know H-5 isn't exploitable because of our router setup" becomes context that flows into the next audit so it doesn't re-surface as Critical.

This is the long-term value of having Cygent in architecture calls: the decisions made in the meeting stop disappearing into a calendar event and start shaping the security behavior of the platform for everything that comes after.

Privacy posture

A few explicit points because they matter:

• Transcription is on whenever Cygent is present. Every word said while Cygent is in the call is transcribed and stored on the Cygent instance. If you don't want a call transcribed, don't invite Cygent.
• Isolated storage. Like every other Cygent data, transcriptions are stored in your isolated instance. Not shared with other tenants, not used to train external models.
• Silent by default. Cygent doesn't respond unless addressed by name. It won't jump into a conversation uninvited.
• Dismissable. Remove Cygent from the Meet at any time (kick it as a participant, tell it "please leave the call" out loud, or ask @cygent leave the meet in Slack).

Limits

• 3 concurrent meetings per instance. Attempting a 4th while 3 are active gets a polite decline.
• 4-hour max duration. Cygent auto-disconnects at the 4-hour mark and posts whatever summary it has at that point.

Related