declared ClaudeBot · Anthropic · not verified
The user agent is ClaudeBot (Anthropic), which identifies itself as an AI agent.
Signals: Known agent user agent
This classification is final for the session. The human pages are not served to agents.
{
"question": {
"slug": "ai-assessment-and-review-interview",
"title": "AI Assessment and the Review Interview",
"difficulty": "medium",
"tags": [
"ml",
"role-ml",
"take-home"
],
"is_free": true,
"locked": false,
"frequency": 0.667,
"last_reported_at": "2026-09-04T22:48:57.178067Z",
"url": "https://www.mokabla.com/en/c/cyshield/ai-assessment-and-review-interview",
"agent": {
"page": "https://www.mokabla.com/agent/en/c/cyshield/ai-assessment-and-review-interview",
"question": "https://www.mokabla.com/agent/_call/get_question?company=cyshield&question=ai-assessment-and-review-interview"
},
"company": {
"slug": "cyshield",
"name_en": "CyShield"
},
"body_md": "## What they send\nAfter the first technical, an AI assessment arrives by email. For the NLP roles it has been three separate tasks in one brief, usually a supervised text task on a dataset they provide, a retrieval or summarisation component, and a small write up of your approach. You get days, not hours, and you submit a notebook or a repository.\n\n## The interview that follows\nThis is the part people are not ready for. The next technical has one agenda: **your submission**. Expect, for every choice you made:\n\n- Why this model and not a simpler one? What was your baseline, and did you beat it?\n- What other approaches could have solved this task, and what is the difference between them? Name two or three alternatives with their trade offs, not one.\n- Why this metric? What does it hide?\n- How did you split the data, and are you sure nothing leaked?\n- What would you change with one more week, and what would you change to put this in production?\n\nThey use the same pattern on your CV. Every project you list gets the same treatment: which model, why that one, do you understand how it works internally.\n\n## Examples\n**Example 1 (strong)** A short README listing the three approaches considered, a table of metrics per approach, one paragraph on why the chosen one won, and a section titled \"what I would do next\". The interview walked through that file and finished early.\n\n**Example 2 (weak)** A notebook with the best score and no history. Asked why not a classical baseline, the candidate had not tried one and could not say whether the transformer was earning its cost.\n\n## Tips\n- Keep every failed experiment in the repository or at least in the README. The alternatives are the interview.\n- State the assumptions you made about the ambiguous parts of the brief instead of quietly picking one.\n- Leave the code runnable end to end. They open it with you on the call.\n- Budget for the wait: the full process has run from one month to four, and the security clearance sits at the end of it.\n",
"follow_ups_md": "- Solve the same task with no deep learning at all. How close do you get?\n- Which of the three tasks did you do worst on, and why?\n- What in your submission would you not put in front of a customer?"
},
"paid": false,
"access": "Question bodies are paid content: a person buys 30 days of access per company (200 EGP or 200 points) in the human app. Agents get titles, tags and difficulty; bodies only for free questions."
}
| action | input | permission | |
|---|---|---|---|
| get_question One question with its body. Bodies are paid content unless the question is free. | company: string, question: string | Allowed | call → |
| list_questions The questions reported at a company: titles, tags, difficulty, whether each is free or paid. | company: string, page: number? | Allowed | call → |
| buy_access Buy 30 days of access to a company. | company: string | Refused R1 This action is for people only | people: https://www.mokabla.com/en/pricing |
| post_comment Comment on a question. | question: string, text: string | Refused R1 This action is for people only | people: https://www.mokabla.com/en/companies |
Allowed actions answer in JSON. Add ?format=json or send Accept: application/json to get this whole page as data.