The GenAI questions freshers are being asked in 2026-27 placements
Four question shapes, the fundamentals that get checked sideways, and why naming a limitation in your project scores higher than claiming it works perfectly.
If your resume mentions an LLM project, GenAI, RAG or "AI/ML", you will be asked about it. What surprises students is how they are asked. Panels are not testing whether you can implement attention from scratch. They are testing whether you understand the thing you put on your resume.
Across the drives we see, the questions land in four shapes.
1. "Explain what you built"
Sixty seconds, the shape of the system. The reliable way to fail this is to open with the tooling — "I used LangChain and Pinecone" — which describes your dependencies rather than your work. Open with the problem instead, then the mechanism, then what a user sees.
2. "Why did you choose that?"
Every choice has an alternative. Why a vector database rather than keyword search? Why that chunk size? Why RAG rather than fine-tuning?
You do not need the best answer. You need to demonstrate you knew there was a choice. "I accepted the default and would tune it by measuring retrieval accuracy" is a solid answer. An invented justification is not, because the follow-up exposes it.
3. "What happens when it goes wrong?"
The question that separates people who built something from people who ran something. What does your system do when retrieval returns nothing relevant? When the user asks about a topic your documents do not cover?
If the answer is "it works fine", you have told the panel you never tested it. Naming a specific failure is the strongest available evidence that the project is yours. The candidates who say "my main failure was questions spanning two documents — retrieval brings back five passages about the fee and none about the deadline" are the ones who get the follow-up conversation.
4. A fundamentals question, dropped in sideways
"What actually is an embedding?" "Is it learning from my chat?" "Why does it make things up?"
These are the ones students fail most, and they are the easiest to fix. The answers, briefly:
- An embedding is a vector representing meaning; similar text lands close together. It is lossy and one-directional — you cannot get the text back from it.
- No, a model does not learn during use. Parameters are fixed at inference; your conversation changes the input, not the model.
- It invents things because it was trained to produce likely text, not true text. A plausible false citation scores well on likelihood. That is the objective working as specified, not a bug.
What this means for how you prepare
Not more features. Depth on what you already have.
Take your existing project and, for each major component, write down one alternative you did not take and one thing that breaks. That exercise takes an evening and moves you past most of the field — because most candidates spend that evening adding a feature nobody will ask about.
We wrote a free six-lesson course on exactly this: how an LLM works at panel depth, what embeddings really are, the RAG diagram you will be asked to draw, and a project scoped so you can defend every part of it. No account needed, and there is a fifteen-question screening set at the end.