Question detection
It answers the question asked of you, not every line
No hotkey, no wake word, no “hey copilot”. Trippi Cue follows the interview and speaks up on one thing: the moment somebody asks you something — and stays quiet the rest of the time, which is most of it.
Trippi Cue
Question
Karl · “…so that is roughly how the team is set up.”
Maya · “And we rotate the on-call weekly.”
Karl · “Anna, how would you handle a release that has to go out during a freeze?”
Suggested answer · asked by Karl
Get it agreed in writing with whoever owns the freeze, ship behind a flag, and keep the rollback one command away. Say who you would ask, not just what you would do.
An illustration of the window, not a screenshot. Two people were talking; only the line addressed to Anna became a card.
What counts as a question
Five shapes, and only one of them has a question mark.
Recognised speech arrives without reliable punctuation, so a question mark is a bonus rather than the signal. These are the shapes that actually get through.
A direct question
“What was the hardest part of that migration?”
The obvious case — and the only one where punctuation helps, when the recognised text has any.
A question without the question mark
“So you would keep the old service running for a week”
Recognised speech often arrives unpunctuated. The opening word carries the signal instead: what, why, how, can, would, do, is, have.
An instruction
“Walk me through how you would design the read path”
Never has a question mark and is still a request for an answer. Imperatives are their own branch: tell me, describe, explain, show us, take us through.
A question with your name in it
“Anna, what would you do differently next time?”
Your name — the first name, the one people actually use in calls — is one of the strongest signals there is.
A follow-up to a follow-up
“And if the traffic doubled?”
Short, contextless on its own, and still yours to answer. The model sees the previous lines, so the thread is not lost.
What it ignores
The quiet is the feature.
A copilot that reacts to everything gets closed in the first ten minutes. Four kinds of line never become a card.
Small talk and agreement — “yeah”, “sure”, “makes sense”, “great”
A line made entirely of these is never a question. This is a large share of what is said in a call, and dropping it on your machine is what keeps the window quiet and the cost sane.
Your own words — Anything you said yourself
Cue does not answer the question you just asked the interviewer — checked twice, in the filter and again when the answer comes back.
Questions aimed at somebody else — “Karl, do you want to take this one?”
In a panel this matters constantly. The model is given the participants and your name, and asked exactly this: was this addressed to the user? If not, no card.
Rhetorical questions — “Who wouldn’t want that?”
A question in form, not in function. This is the judgement a keyword list cannot make, which is why the decision belongs to a model and not to a regular expression.
How the decision is made
Three stages, two of them filters.
-
1
Local filter, on your machine
Runs on every recognised line, costs nothing and sends nothing anywhere. Deliberately generous: it throws away the obviously empty, not the doubtful. A missed question costs more than an extra check.
-
2
The model decides ownership
The recent lines, the participant names and your own name go off as one prompt. The model answers whether the closing turn was addressed to you, what type it was, who asked, and how sure it is.
-
3
The gate on the way back
Below the confidence threshold — nothing is shown. Same question again within a minute — merged into the existing card instead of a second one. Automatic answers switched off — only your own questions get through.
When it gets it wrong
What to do about a false positive — or a miss.
Both happen. Neither needs support: every fix here is one click inside the window.
It answered something not meant for me
Dismiss the card — one click, and the window is clean again. If it keeps happening on a panel, the likeliest cause is that your name never appears in the call: set it in the window so the model knows which participant you are.
It stayed silent when I was asked
Ask it yourself in the field at the bottom: the same card, on demand, with the meeting still as context. That path skips every filter on purpose.
It answers too often
Turn off automatic answers in the settings. The window then reacts only to what you type — some people run it that way from the start.
It heard the question wrong
The strip of recognised lines under the cards shows what it actually heard. A name mangled by recognition is the usual culprit, and it is visible right there.
Timing
Why it arrives when it does.
A card lands a few seconds after the question is finished — the time it takes to recognise the closing words and get an answer back. It is not instant, and it does not need to be: in an interview, a pause of a couple of seconds before you answer is normal, and an answer that arrives inside it is on time.
Waiting for the turn to end is deliberate. Answering half a question produces an answer to something nobody asked, and the correction costs more attention than the wait.
Privacy
What detection sends, and where.
Detection itself starts on your machine: the local filter reads the line and sends nothing. Only what passes it — a dozen recent lines, the participant names, your name — goes to our backend and on to the model. That window and the answer are kept for 30 days and then deleted; the audio never reaches us at all.
FAQ
About detection, specifically.
How long does detection take?+
A card appears seconds after the turn ends. The delay is mostly recognition and the model call — there is no polling interval and no batching to wait for.
Does it need a hotkey or a wake word?+
No. There is nothing to press and nothing to say: the last click you make is the one that starts the session.
What if two people ask me things at once?+
Each request gets its own card. Three are on screen at a time; the fourth pushes out the oldest.
Does it work when the interview is not in English?+
Recognition follows the language of the call, and the lines also arrive translated into yours. The local filter’s word lists are English, so in other languages the model does more of the work — which means slightly more calls, not fewer cards.
Can it tell a task from a question?+
Yes, and it says which: a question gets an answer you can say out loud, a task gets how to approach it rather than what to say.
Why is confidence not shown on the card?+
Because a number next to an answer invites you to argue with it instead of reading it. It is used as a threshold: below it, the card never appears at all.
Does it store the questions it detected?+
The recognised lines and the answer are kept for 30 days, so a poor answer can be told from a poor question, and are then deleted automatically. The audio is never stored at all — it does not reach our servers in the first place.
The next question aimed at you, caught while it still matters.
Trippi Cue is in the Chrome Web Store, and starting is free. Add it before your next interview.