The problem
First-round interviews eat a panel’s day: the same opening questions, candidate after candidate. Today’s AI screening tools are mostly recorded video or chat, so nobody can step in mid-conversation, and panels end up re-interviewing the good candidates anyway.
What Ally does
Ally joins a normal Google Meet as a participant and runs the interview by voice, with follow-ups based on the answers. The panel keeps control:
- Hand up: a panelist raises their hand, and Ally stops mid-word, says “Sure, go ahead”, and keeps transcribing.
- Hand down: once every panel hand is down and the room is quiet, Ally picks up where it left off, knowing what the panel just asked.
- After the call: a scored report against the rubric, with evidence quoted from the transcript.
Take over, hand back, skip and end are always available from a control page, or by typing ally wait in the Meet chat.
The hard part
Meet has no API for raised hands. Ally’s patch to the Attendee bot watches the participant tiles inside the bot’s headless Chrome and streams hand_raise / hand_lower events over a websocket, with a slower webhook copy de-duplicated by event id.
The silence check exists because Meet lowers a hand automatically as soon as its owner starts talking. Resuming on “hand down” alone would talk over the panelist.
Measured on a live call
| Step | Time |
|---|---|
| Hand raised → Ally silent | ~0.23 s |
| Claude first token | 1.1–1.5 s |
| Cartesia first audio | ~0.08 s |
| Hand lowered → Ally resumes | ~6 s (5 s silence check + 1.1 s) |
What’s next
- Hindi and Hinglish interviews
- Microsoft Teams (same patch, new selectors)
- ATS integrations that push reports automatically
- A call-assist mode for sales and support, where a supervisor raises a hand to take over