Who runs this: the MeetAssist team. MeetAssist makes an AI interview assistant too, so we publish the inputs, the recordings and the raw numbers. If our method favours anyone, you can see it and test it yourself.
What the test measures
- Speed: how long after the interviewer stops talking the first words of an answer appear, and when the answer finishes on screen.
- Follow-ups: whether the assistant keeps the context when the next question says “in this case” or “the same problem”.
- Use of your documents: whether it uses the facts from an uploaded CV and stories, or makes things up.
- Technical and coding quality: whether explanations are correct, and whether the code actually runs. We test code with automated tests; we do not judge it by eye.
It is a small hands-on test, not a ranking. One recording on one machine is an observation; we say so in every review.
The candidate and the files
The candidate, Alex Morgan, is fictional: a senior full-stack engineer with 8 years of experience. The CV and the stories document are written so we can check answers against them: exact numbers (a billing rollout of 10% → 25% → 50% → 100%, failure rate 2.0% → 1.3%), a specific React bug, and a list of things the AI should not know.
- Candidate CV (PDF): upload it as the CV.
- Candidate stories (DOCX): upload it as a reference document.
The questions
The questions are pre-recorded with the same AI voice for every product and played from our test page. Each follow-up depends on the answer before it. The tester starts the next question only after the answer has finished. The coding task is shown on screen and read with the product's screenshot feature.
| Step | Question | Audio |
|---|---|---|
| Behavioral | Tell me about a time you had a disagreement with a product manager. How did you handle the situation? | MP3 |
| Follow-up | Why did you choose that particular approach instead of the alternatives? | MP3 |
| Deeper | And what if that hadn't worked? What would you have done differently? | MP3 |
| New question (no story for it) | Tell me about a time you failed to meet a deadline. What happened? | MP3 |
| Question in Spanish | Háblame de un proyecto del que estés orgulloso. ¿Cuál fue tu papel? | MP3 |
| Technical | Imagine we have a React application with an editable list of tasks. Each task has a text input, and we're using array indexes as React keys. When we insert a new task at the top, some input values appear under the wrong tasks. What's causing this? | MP3 |
| Follow-up | Okay, but what exactly is React reusing in this case? | MP3 |
| Deeper | And what if they don't have stable IDs? How would you fix it without losing what the user already typed? | MP3 |
| Coding | Trapping Rain Water (LeetCode 42, Hard), shown on screen with an empty JavaScript function. | Task page |
| Follow-up | Can you solve the same problem in O(n) time and O(1) extra space? Explain your approach. | MP3 |
| Deeper | Why is it safe to move the pointer on the shorter side? | MP3 |
Two questions test honesty and languages. The candidate's documents contain no story about a missed deadline, so a good assistant should say it has no example rather than invent one. The Spanish question checks whether the assistant hears and answers another language.
We run each product with the CV and stories and without any documents. A review says which runs it is based on.
Earlier version. Our first Parakeet run (October 15, 2026) used different wording for the two behavioral follow-ups (MP3, MP3), had no new or Spanish question, and played each next question 10 seconds after the previous one.
How we time the answers
- Start: the moment the interviewer's voice stops. We find it in the recording's sound track by lining up the original audio file with what the microphone heard (to about ±0.01 s). For screenshot tasks the start is the product's reaction to the screenshot action.
- First answer: the first video frame where words of the answer (not just an empty panel or a loading sign) are visible.
- Visible completion: the first frame after which the answer on screen stops changing. If the answer is longer than the window, we say so: the end of such answers is not visible, so that time is a lower bound.
- Steps in between: when a product shows its own stages (for example “Checking…” or “Question detected”), we time those too.
- From the trigger: we also time each answer from the moment the app shows it is answering (“Question detected”, a screenshot, or our hotkey press in apps without automatic answers) to the first words and to the full answer. This lets us compare apps with and without automatic answers. Apps with automatic answers may do part of the work before that moment, so their times from the trigger look shorter.
We film the screen with a phone, because these apps do not show up in normal screen recordings. Precision depends on the recording: ±0.03 s at 30 frames per second, ±0.017 s at 60. The timer on our test page starts when the browser reports that the clip ended, which is about 0.2 s after the voice actually stops, so numbers read off that timer in screenshots are slightly lower than our measured times.
How we score the answers
Each answer: 10 points.
| Criterion | Points | What we check |
|---|---|---|
| Correct | 0–4 | What it says is true: technically, and about the candidate where the documents have the facts. |
| On topic | 0–2 | It answers the question that was asked and keeps the context of follow-ups. |
| Your facts | 0–2 | It uses the candidate's real facts. A made-up story or detail with no label (such as “example”) gets 1; contradicting the documents gets 0. |
| Usable live | 0–2 | You can read it and say it during the call: clear, and it fits the window. |
No answer gets 0. A question in a language other than the one set in the app is not scored; we describe what happened instead. Code is copied from the answer exactly and run against automated checks: the task's examples, tricky cases (empty input, one item, all values equal), thousands of random inputs compared with a simple, slow solution that is certainly correct, and one very large input.
The scorecard: three categories, each out of 10.
| Category | How we calculate it |
|---|---|
| Answers | The average score of all scored answers, with and without documents. |
| Speed | From the typical wait for the first words after the question ends: 12 − 2 × seconds (1 s = 10, 2 s = 8, 3 s = 6, 4 s = 4, 6 s or more = 0). |
| Coding | Code works on every check = 6, best-known complexity = 2, each follow-up explanation right = 1 (two follow-ups). |
Overall is the average of the three. Next to it we list what we found but do not score, such as made-up stories without a label and questions in another language. The same rules apply to every product.
How we time the answers, in three parts. Detect: from the end of the question until the assistant starts answering (with automatic answers, until it decides it heard a question; with manual tools, until our hotkey press). Think: until the first words appear. Write: until the whole answer is on screen.
Results and raw data
Parakeet AI
Review: Parakeet AI review. Same settings in every run: Gemini 3.5 Flash Lite, automatic answers on, script + bullets, balanced length, simple tone, call language English.
October 16, 2026, with CV and stories (11 questions)
- Timings and scores (CSV)
- Every answer, word for word (TXT)
- Video of both October 16 runs: YouTube
October 16, 2026, without documents (5 behavioral questions)
October 15, 2026, with CV and stories (9 questions, earlier version of the questions)
- Timings and scores (CSV)
- Every answer, word for word (TXT)
- The coding answer and its tests (JS): run it with
node parakeet-ai-2026-10-15-trap.test.js. The October 16 code passes the same tests. - Video: YouTube Short
Known limits
- A few recordings per product: good for seeing how a product behaves, not for statistics.
- Speed depends on the network, the region and the AI model selected. We record the model when the product shows it.
- Questions are played through speakers, as a real call would be. A product that captures audio differently may behave differently with headphones.
- We use only features available to a normal user. We do not test claims about being hidden from screen sharing or other software.
Or add the Chrome extension