An AI internship viva tests whether you understand your own work: where the data came from, why you chose the model, what the numbers really mean and what could go wrong. This question bank groups the likely questions into data, model, metrics and ethics, and gives each an answer outline you can adapt to your project.
How an AI Internship Viva Differs From the Written Report
A report lets you polish sentences. A viva removes that cushion. An examiner can interrupt, change direction and ask “why” three times in a row. The aim is rarely to catch you out. It is to separate a student who ran a notebook from a student who understood it.
For a short project, such as a 14-day internship, examiners usually expect a narrow scope done honestly rather than an ambitious scope done vaguely. A small project that you can fully explain beats a large one you cannot. Preparing for the viva means preparing to defend decisions, limits and trade-offs.
A useful habit: for every slide or paragraph in your report, write one sentence beginning “I chose this because…” and one beginning “The weakness of this is…”. Those two sentences answer most follow-up questions before they are asked.
Before You Start: Build a One-Page Defence Sheet
Fill in this sheet before you open the question bank. If you cannot complete a line, that is where the examiner will probe.
- Problem statement: one sentence, with the prediction target named.
- Data: source, number of rows, number of features, time period, licence or permission to use it.
- Cleaning steps: what you removed, imputed or transformed, and why.
- Baseline: the simplest comparison you used (majority class, mean prediction, a rule).
- Model: name, key settings, why it suits this data size and type.
- Metric: the one you report first and the reason.
- Limits: two honest weaknesses.
- Next step: one improvement you would make with more time.
Free 2-Week CareerSwami Internship
Online, 14 days, with a free offer letter and a free completion certificate once you finish the work. Check with your college that it fits your requirement before you enrol.
Data Questions: Source, Quality, Leakage and Splits
Examiners start here because every later result depends on it. Weak data answers undermine a good model.
| Likely question | Answer outline |
|---|---|
| Where did your data come from, and were you allowed to use it? | Name the source, the date accessed and the licence or terms. If it was a public dataset, say so. If it was scraped or supplied, state what permission existed. Never leave this vague. |
| How many records and features, and what does one row represent? | Give exact counts from your notebook and define the unit of observation (one customer, one image, one review). |
| How did you handle missing values? | State the share missing per column, the method (drop, mean, median, mode, model-based) and why. Admit if imputation could bias results. |
| How did you split the data? | Give the ratio, whether it was random, stratified or time-based, and the random seed. Explain why the test set was untouched until the end. |
| What is data leakage, and did you check for it? | Leakage means information from outside the training context, such as the target or future data, enters the features. Describe one check: scaling fitted only on training data, or dropping columns derived from the target. |
| Is your dataset balanced? | Report the class proportions. If imbalanced, explain what you did (class weights, resampling, a different metric) or why you left it. |
| Is the data representative of real use? | Compare the data’s population and time period with the intended users. Name one group or situation it may not cover. |
| Why did you choose these features? | Link each feature group to the problem. Mention any you dropped for correlation, irrelevance or leakage. |
A Tip for Answering Data Questions
Lead with numbers. “About 8% of the income column was missing, and I filled it with the median” sounds like someone who looked at the data. “I cleaned the data” does not.
Model Questions: Choice, Training, Tuning and Overfitting
Here the examiner wants reasoning, not a list of algorithms. Saying “I used Random Forest because it is popular” invites trouble. Saying “I used it because the data was tabular, mid-sized and had mixed feature types, and I compared it with logistic regression” does not.
| Likely question | Answer outline |
|---|---|
| Why this model and not another? | Tie the choice to data type, size, need for interpretability and training time. Mention at least one alternative you tried or considered. |
| What was your baseline? | Describe the simple reference (for example always predicting the majority class) and the score it achieved. A model must beat this to matter. |
| Explain how the model works in two minutes. | Use plain language and one small picture or analogy. Avoid reciting formulas you cannot derive. |
| What is overfitting, and how do you know whether you had it? | Overfitting is strong training performance with weaker unseen performance. Quote the training and validation scores and the gap between them. |
| How did you tune hyperparameters? | Name the parameters, the search method (manual, grid, random), and that tuning used validation data or cross-validation, not the test set. |
| What is cross-validation and why use it? | Repeated train-validation splits give a steadier estimate when data is small. State the number of folds if you used it. |
| What would you change with more data or time? | Pick realistic items: more data, better features, error analysis, a different model family. Avoid promising a deep network by default. |
| Can you explain one prediction? | Use feature importance, coefficients or an example row. Say clearly that importance shows association, not cause. |
| Did you use a pre-trained model or library code? | Say exactly what was reused and what you wrote or adapted. Honesty here protects you. |
Metrics Questions: Accuracy, Precision, Recall and Error Analysis
Metrics questions reveal quickly whether you understand the problem’s costs. Know the definitions cold and be able to say which error hurts more in your use case.
| Metric | Plain meaning | When it misleads |
|---|---|---|
| Accuracy | Share of all predictions that were correct | Imbalanced classes: 95% accuracy can mean the model always predicts the majority |
| Precision | Of the items predicted positive, how many truly were | Ignores positives the model missed |
| Recall | Of the truly positive items, how many were found | Can be pushed up by over-predicting positives |
| F1 score | Balance of precision and recall | Hides which of the two is weak |
| RMSE / MAE | Typical size of numeric prediction error | RMSE penalises large errors heavily; MAE does not |
| ROC-AUC | Ability to rank positives above negatives across thresholds | Can look good on highly imbalanced data |
Metric Questions With Answer Outlines
- Why did you pick this metric? Connect it to the cost of mistakes. In a screening task, missing a true case (low recall) usually costs more than a false alarm.
- What does your confusion matrix show? Read it aloud: true positives, false positives, false negatives, true negatives, and which cell concerns you most.
- Is 92% accuracy good? Answer “compared with what?” Give the baseline score and the class balance before judging.
- How did you choose the decision threshold? State the default 0.5 if you kept it, or the reason for changing it and what it did to precision and recall.
- Where does the model fail? Show two or three misclassified examples and a pattern, such as short texts or rare categories.
- How stable is the result? Mention the seed, repeated runs or cross-validation spread. Do not present one lucky run as the truth.
Ethics and Responsible Use Questions
Even a small project can attract these questions, and a thoughtful answer often lifts the whole viva. Keep answers concrete and tied to your data.
| Likely question | Answer outline |
|---|---|
| Could your model be biased? | Name a sensitive attribute or proxy in your data (location, gender, language). Say whether you checked performance across groups, or that you did not and would do so next. |
| Does your data contain personal information? | State whether it was anonymised or public. Explain that you removed identifiers and did not try to re-identify anyone. |
| Who could be harmed by a wrong prediction? | Describe the affected person and the consequence, then say a human should review high-stakes outputs. |
| Should this model be deployed? | A fair answer is “not yet, as a prototype”, with the checks required first: wider data, group-wise testing, monitoring. |
| Did you use generative AI tools for the work or report? | Answer truthfully and specifically: what the tool helped with and what you verified yourself. Be ready to explain any code or text you submitted. |
| Can the model’s decision be explained to a non-technical user? | Give a short plain-language explanation and note any limits of interpretability. |
Worked Example: Defending a Loan-Default Classifier (Illustrative)
The following is an invented scenario to show answer structure. The figures are made up for illustration; no code was run for this article.
Project: a student predicts whether a loan applicant will default using a public tabular dataset of 10,000 rows. About 10% of rows are defaults.
Examiner: “Your accuracy is 91%. Is that good?”
Weak answer: “Yes, it is above 90%, so the model works well.”
Strong answer outline:
- Reframe: “Only about 10% of the data are defaults, so a model that always predicts no default would score roughly 90%. My baseline is therefore about 90%, and 91% is only a small gain.”
- Switch metric: “So I looked at recall for the default class, which was 38%. The model misses most defaulters.”
- State the cost: “A missed default is costlier than a rejected good applicant, so recall matters more here.”
- Show action: “I tried class weights, which raised recall to 61% while precision fell to 40%. That trade-off is the honest picture.”
- Acknowledge limits: “I have not tested performance across income groups, so I would not claim it is fair yet.”
Notice the pattern: restate the question in terms of the problem, bring in the baseline, give the better metric, own the trade-off, and name a limit. That five-step shape works for most metric questions.
Manav Rachna University: Preparing for the Viva as an MRU Student
Start by making sure you are reading the right pages. The shared website manavrachna.edu.in hosts both Manav Rachna University and a separate institution, Manav Rachna International Institute of Research and Studies. The site lists them as separate institutions, each with its own section, so an internship rule you find under the MRIIRS section does not automatically apply to you. The MRU home page describes the university as a State Private University recognised by UGC, and that is the section to treat as yours.
The same caution applies to examples. The MRIIRS B.Tech ECE page, for instance, lists an Industrial Internship / Minor Project without giving a duration, credits or evaluation rules, and it is not an MRU page. No MRU regulation on viva format, marking, online internships or required documents has been verified for this article, so do not assume your department’s expectations from this guide or from another institution’s page.
MRU’s site also links a B.Tech Computer Science Engineering programme page and an Act, Statutory and Ordinances page from its home page navigation. These are the places to start looking for programme structure and university-wide rules, though their contents were not verified here. If your internship is a free online one, such as the free two-week internship from CareerSwami, your viva may be an internal presentation in which your department decides what counts. That is why written confirmation matters.
Verification Steps for Your MRU AI Internship Viva
- Ask your internship coordinator or faculty mentor in writing whether a viva is required, who examines it and how long it lasts.
- Request the current internship or project guidelines for your programme and semester, and check the document’s year.
- Ask whether an online internship is acceptable for your requirement and what proof (offer letter, completion certificate, report) the department wants.
- Check with your school office whether the viva is marked, and whether it is internal only or involves an external examiner.
- Confirm the report format and submission deadline, and whether slides or a demo are expected.
- Look in the MRU Act, Statutes and Ordinances page and your programme page for any rules on training or projects, and ask your coordinator to confirm how they apply to you.
A 14-Day Plan to Get Viva-Ready During a Two-Week Internship
| Days | Focus | Viva preparation task |
|---|---|---|
| 1-2 | Problem and data | Write the problem sentence and the data section of the defence sheet |
| 3-5 | Cleaning and baseline | Record missing-value rates, split method and baseline score as you go |
| 6-8 | Modelling | Note why each model was tried and what the validation scores were |
| 9-10 | Metrics and errors | Build the confusion matrix and collect three failure examples |
| 11 | Ethics and limits | Write answers on bias, privacy and deployment |
| 12 | Slides | Keep to a short deck; one idea per slide |
| 13 | Mock viva | Ask a friend to ask “why” after every answer; time yourself |
| 14 | Final polish | Re-run results, confirm numbers match your slides and report |
Delivery Techniques for Answering Under Pressure
- Pause before answering. Two seconds of silence reads as thought, not weakness.
- Answer first, explain second. Start with the direct answer, then give the reason.
- Say “I do not know, but here is how I would find out.” It is far better than inventing an answer.
- Keep numbers consistent. If a slide says 0.84 and you say 0.86, the examiner will notice.
- Point to evidence. Open the confusion matrix or the plot rather than describing it from memory.
- Do not oversell. Use “suggests” and “in this dataset” rather than “proves”.
Warning Signs That Your Answers Are Not Ready
- You cannot state the baseline score without looking it up.
- You describe the model only by its name and not by what it does.
- You report accuracy alone on imbalanced data.
- You tuned on the test set, or cannot say whether you did.
- You cannot explain a library function you called.
- Your ethics answer is only “there are no ethical issues”.
Frequently Asked Questions
How many viva questions should I prepare for an AI internship?
Quality matters more than volume. Prepare thorough answers for around four to six questions in each of data, model, metrics and ethics, and make sure you can tie each to your own numbers and decisions.
What if the examiner asks about an algorithm I did not use?
Say you did not use it, then explain briefly what you know and why you chose your approach instead. Comparing it with your model on data size or interpretability shows understanding without pretending experience.
Do I need to memorise formulas for precision, recall and F1?
Know the definitions in words and be able to compute them from a confusion matrix. Reciting the formula matters less than explaining what a high or low value means for your problem.
How should I answer if my model performed poorly?
Present it honestly. Show the baseline, your error analysis and your hypotheses about why it underperformed. Examiners usually respect a clear diagnosis of a weak result more than an unexplained strong one.
Can I bring a demo or notebook to the viva?
That depends on your department’s format, so ask your coordinator in advance. If allowed, keep the notebook clean and ordered, and have screenshots ready in case of technical failure.
Is it acceptable to say I used ChatGPT or other AI tools?
Whether a tool is permitted is set by your department, so check the rule. If you did use one, say so plainly and be ready to explain every line of code and sentence you submitted as your own work.
How do I explain data leakage simply?
Say it is when the model gets information during training that it would not have in real use, such as a column that is calculated from the answer. It makes results look better than they will be in practice.
How can a two-week online internship project be defended convincingly?
Keep the scope narrow, document every decision as you go and rehearse a mock viva. A small, well-understood project with a stated baseline and honest limits is easier to defend than a broad one.
Related Artificial Intelligence guides
- Gokhale Institute Students: Applying AI to Economics Data in a Short Internship
- Nrupathunga University Students: How to Pick Between Two AI Internship Offers Using a Weighted Scorecard
- Shaheed Bhagat Singh State University Students: A Prompt Engineering Log for Your GenAI Internship
- All Artificial Intelligence internship guides
Ready to put this into practice?
Join the free two-week CareerSwami online internship in Artificial Intelligence. The internship, the offer letter and the completion certificate are free.
