AI interview scenarios,
twice a day.
One scenario-based question of the kind technical AI interview panels actually ask — from “what does a transformer compute?” to “which architecture fits a bank automating its finance close?” The question goes up at night. The answer, and what the panel listens for, follows 36 hours later. Every explanation cites the sources it rests on.
How one question runs
The same rhythm every day, so the question reaches one side of the world while the answer reaches the other. Times are India Standard Time (UTC+5:30).
A scenario and four options
A short video on YouTube and Instagram, and a poll on LinkedIn. No hints, no answer in the caption, nothing that gives it away. You commit to an option.
The poll stays open
Voting runs its course before anything is revealed. No answer appears on any channel while the poll is still open.
Why, and what the panel wanted
The correct option, why each of the other three falls short, what a panel is really listening for, and the sources behind every claim.
What a question looks like
A sample, shown with its answer. On the channels, this half arrives 36 hours after the question.
Your retrieval-augmented support assistant answers from 40,000 policy documents that change weekly. It runs CPU-only, with a 400 ms first-token budget. Today it retrieves 50 chunks and sends all of them to the model.
CPU-only, 400 ms first-token budget, 40,000 documents changing weekly. What do you change first?
- ASwitch to a larger embedding model for better recall
- BAdd a lightweight reranker and pass only the top 5 chunks
- CFine-tune the language model on the policy documents
- DRaise the context window and send all 50 chunks
Answer — B
Prompt length drives both time-to-first-token and cost. A small cross-encoder reranker over the 50 candidates costs a few milliseconds on CPU and lets you send 5 well-chosen chunks instead of 50, which is what brings first-token latency back under budget.
On the channels this slide carries the linked sources each number and quote came from.
Our standard
A daily question is only worth answering if the answer holds up. So the bar is set before anything is written.
Every claim traced to a source
Quotes must appear verbatim in the paper, documentation or report they come from, and every number must come from that source or from arithmetic shown in full. Anything that fails is dropped, not softened.
Reviewed before it ships
Each question is read back against its own sources before it is published, and read cold once more for anything that gives the answer away. Questions that do not survive that are not posted.
Built from real decisions
The scenarios mirror trade-offs teams actually face: a memory budget that will not fit, a first-token deadline, a cost per request, an agent that can take an action it cannot undo.
No career promises
These are practice scenarios and explanations, not hiring advice, and not a promise about any job, employer or salary.
Where it posts
The same question on every channel, in the format that channel does best.