Common Capstone Defense Questions for AI Projects (Answers)

AI capstone defense panels in 2026 ask harder questions than they did in 2022, because they have seen 4 years of AI projects and learned which student demos are “real” vs “GPT wrapper.” This guide is 40+ panel-tested questions across 6 categories, each with a sample answer template you can adapt.

If you can answer all 40 confidently, you will pass. If you stumble on 5+, study those before defense day.

Common Capstone Defense Questions for AI Projects (Sample Answers)

📌 Defense-prep strategy: Print this page. Highlight every question you cannot answer in 1 minute without notes. Practice those out loud 3x. The panel will pick from these 40, and the ones you cannot smoothly answer are the ones they will dig into.

Category 1: Model & Architecture (10 questions)

Why did you choose this specific model (e.g. YOLOv8, BERT, GPT-4o-mini)?
Sample answer: “We evaluated 3 candidates (X, Y, Z) on our test set. Model A scored 89% accuracy vs Y at 84% and Z at 80%. A also had the smallest model size (45MB) for our mobile deployment constraint. Trade-off: slower inference (180ms vs 90ms) which is acceptable for our use case.”
What is the accuracy of your model?
Sample: Quote precision, recall, F1 separately. NEVER just “accuracy.” Panels know that imbalanced datasets make accuracy misleading.
What is your dataset size and source?
Sample: “1,247 images, 70% training (873), 15% validation (187), 15% test (187). Sourced from Kaggle dataset X (open license), supplemented with 200 locally captured samples for domain matching.”
How did you split train/validation/test?
What loss function and optimizer did you use? Why?
What is your model’s F1 score? Precision? Recall?
How did you handle class imbalance?
Why this architecture instead of a simpler one (logistic regression baseline)?
Sample: “Our baseline logistic regression achieved 71% F1. CNN reached 89%. The 18 percentage point gain justifies the added complexity for our use case.”
What is overfitting? How did you detect and prevent it?
Could you defend the result without your computer?

Category 2: Data & Ethics (8 questions)

Where did your training data come from?
What consent / data privacy considerations apply? (Data Privacy Act of 2012 if you collected PH data)
What biases might exist in your dataset?
Sample: “Our face dataset is 85% lighter-skinned. We expect lower accuracy on darker skin tones, a known fairness issue in face recognition. For deployment we would need to collect additional samples and re-train.”
How did you handle missing or noisy data?
Did you anonymize personal data?
What happens if your model makes a mistake in production?
How do you ensure your model isn’t harmful or discriminatory?
Are you compliant with Republic Act 10173 (Data Privacy Act)?

Category 3: Implementation Detail (10 questions)

Walk us through your code on this slide. (Have your repo open, highlight the 20-line core function)
How long does inference take per query?
How much memory / GPU does your model need?
What framework did you use (PyTorch, TensorFlow, scikit-learn) and why?
How do you handle API rate limits / OpenAI costs?
Show us your model training loop.
How did you tune hyperparameters?
What was the hardest bug you fixed? (Memorize a real story; panels love this)
How do you log and monitor production errors?
What happens if your training data and production data drift apart?

Category 4: Deployment & Scaling (6 questions)

Where will this be deployed (cloud, on-prem, mobile)?
What is the cost per inference at scale?
How does the model update when new data comes in?
What happens if 1000 users hit the API simultaneously?
How do you handle versioning of the model in production?
What’s your backup plan if the model service goes down?

Category 5: Business / Real-World Impact (4 questions)

Who is your target user and what problem does this solve?
Do you have a real client / endorsement letter?
What is the cost-benefit analysis? (“System saves 4 hours/day of manual work, paying for itself in 3 months”)
How is this different from existing commercial solutions?

Category 6: Future Work (2 questions)

What would you improve given another semester?
How would you extend this to multi-language / multi-domain / mobile?

Frequently Asked Questions

How long should my answer be for each defense question?

30-90 seconds. Long enough to show depth, short enough that the panel doesn’t lose patience. Practice with a timer; if your answer goes over 2 minutes, cut it down.

What if I get a question I genuinely don’t know the answer to?

Say so honestly: ‘I’m not 100% sure but my best understanding is…’ or ‘That’s a great question, I’d need to verify, but my initial thought is…’. Never bluff. Panels respect honesty far more than guesses presented as facts.

How can I prepare if my AI capstone uses a third-party API (OpenAI, Anthropic)?

Be ready to defend: (1) Why this provider not the others, (2) Cost projection at scale, (3) What happens if the API changes or shuts down, (4) Data privacy implications of sending data to a third party. These are the most common gotcha questions for API-based AI capstones.

Do I need to show the code during defense?

Have it ready and open in case the panel asks. Most panels don’t read code line-by-line but may ask you to walk through the core function. Pre-select 2-3 ‘showcase’ functions to walk through if asked.

What evaluation metrics should I report for an AI capstone?

Always include: accuracy, precision, recall, F1 (for classification) or MAE/RMSE (for regression) or BLEU/ROUGE (for NLG). Show a confusion matrix and a few qualitative examples of correct vs incorrect predictions. Avoid reporting only training accuracy without test accuracy.

What’s the most common reason capstones fail defense?

Inability to explain WHY a design choice was made. Panels can tell when a student copy-pasted code without understanding it. Practice articulating the reasoning behind every architecture decision, even if the answer is ‘I chose X because the tutorial used X.’