What panels actually ask about AI
The four question shapes that come up, why "I used the OpenAI API" is a losing answer, and the standard you are being measured against.
If you have "GenAI" or an LLM project on your resume in 2026, you will be asked about it. Not by a research scientist — by an engineer with a hiring slot to fill and twenty minutes. What they are testing is narrow and predictable, and it is not what most students prepare for.
The four shapes
Nearly every question a fresher gets about GenAI is one of these four.
1. "Explain what you built." They want the shape of the system in sixty seconds. Where does the data come from, what happens to it, what does the user see. Students fail this by starting with the framework — "so I used LangChain and Pinecone" — instead of the problem.
2. "Why did you choose that?" The trap question. Every choice you made has an alternative, and they want to know you were aware of one. "Why a vector database and not a keyword search?" "Why chunk at 500 tokens?" An honest "I followed a tutorial and that was the default" scores higher than an invented reason, but not much higher. Knowing one real trade-off scores far higher than both.
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 the retrieved documents are irrelevant? When the model makes something up? When the user asks about a topic you have no documents for? If your answer is "it works fine", you have told them you never tested it.
4. A fundamentals question, dropped in sideways. "What actually is an embedding?" "Is the model learning from my documents?" These are checking whether you understand the machinery or just the API surface. This is the one students most often fail, and the easiest to fix — it is the subject of the next three lessons.
Why "I used the OpenAI API" loses
Because it is a description of a dependency, not of your work. Consider two answers to the same question:
A. "I built a chatbot for college FAQs using the OpenAI API and LangChain."
B. "College FAQs were scattered across six PDFs, so a student's question rarely matched any single document's wording. I split the PDFs into passages, embedded them, and at query time retrieved the five closest passages and passed them to the model as context. The failure I could not fix was questions that span two documents — retrieval brings back five passages about hostel fees and none about the deadline, so the answer is confidently half-right."
Answer B is not longer because it is padded. It names the problem, the mechanism, and a real limitation. It also hands the interviewer their next question, which is exactly what you want — you are now being interviewed about something you understand.
Notice that B never claims the project was good. Naming a limitation is not a weakness in an interview; it is the strongest available signal that you actually built the thing. Panels see dozens of projects that allegedly work perfectly.
The standard you are being measured against
A fresher is not expected to know how attention is implemented. You are expected to:
- describe your system's data flow without referring to a library name
- know one alternative to each major choice, and why you did not take it
- name one thing that breaks, specifically
- answer a fundamentals question without bluffing
That is the whole bar. It is reachable in a week, and most candidates do not clear it — not because it is hard, but because they spent that week adding a feature instead.
What the rest of this course does
The next three lessons cover the fundamentals that get asked: how an LLM produces text, what an embedding is, and how retrieval is actually assembled. Lesson 5 covers prompting without the folklore. Lesson 6 walks through a project you can build and defend, including the four follow-up questions it will attract.
There is a free practice set at the end. It is fifteen questions, and it is the same style of question a screening round uses — no account needed.