AI and machine learning final year project mentoring

Why do many AI/ML projects struggle at the viva?

Machine learning projects are the most popular final year choice and the easiest to get wrong. The common pattern is a model copied from a tutorial, trained on a famous dataset, reporting 98% accuracy with no explanation of what was done to the data. Examiners have seen hundreds of these. Their questions are predictable: where did the data come from, how did you split it, why accuracy and not recall, what does the model get wrong, and how would it behave on new data.

A strong project answers those questions before they are asked. That is what mentoring focuses on: a sensible dataset, a baseline, honest evaluation and an explanation you understand.

How do you choose a dataset?

Use a dataset you are allowed to use and can describe. Public sources such as Kaggle, the UCI repository, government open data portals and research datasets are fine if you cite them. Data you collect yourself, such as survey responses, photographs from a local farm or readings from a sensor, makes a project more original, but plan the collection effort early.

Check the size, the class balance, missing values and whether the labels are reliable. A dataset of 300 rows can support a solid classical ML project; it cannot support a deep network. Your mentor reviews the dataset choice before you write a line of modelling code.

What does a sound modelling workflow look like?

1. Explore the data: distributions, missing values, outliers, class balance.
2. Split into training, validation and test sets before any tuning, and never touch the test set until the end.
3. Build a simple baseline such as logistic regression, a decision tree or a pretrained model with minimal changes.
4. Improve step by step: features, model choice, hyperparameters. Record each result in a table.
5. Explain the model with feature importance, SHAP values or saliency maps.
6. Wrap the final model in a small demo, for example Streamlit, Flask or a simple mobile front end.
7. Write up results honestly, including where the model fails.

Which AI/ML topics are worth considering?

• Crop disease detection from leaf images collected locally, with a mobile demo
• Early warning of student dropout risk with explainable predictions for counsellors
• Air quality forecasting for an Indian city from public monitoring station data
• Resume screening that flags bias in its own scores
• Sign language alphabet recognition with a webcam
• Energy demand forecasting for a campus or housing society

How does mentoring work for an AI/ML project?

1. Send your topic idea or interests, the data you have access to and your deadline.
2. Free scope call to confirm the problem, the dataset and the evaluation metric.
3. Pay 50% to book. The balance is due after you see the work reach the agreed stage.
4. Notebook reviews every 2 to 3 days: data cleaning, baseline, experiments.
5. Demo guidance so the model runs reliably on the day.
6. Report and PPT checks, with results tables built from your own runs.
7. Mock viva on data, features, metrics and limitations.

What do you get?

• Dataset and metric review before modelling starts
• Notebook reviews with explanations of every change
• An experiment log you can turn into your results chapter
• Guidance on a simple, reliable demo
• Report and PPT formatting help
• Mock viva focused on ML questions

Which plan fits?

How should the results chapter be written?

Start with a table of every experiment you ran: model, features, key parameters and the metric on the validation set. Then report the final model once on the test set. Add one or two plots that matter, such as a confusion matrix or predicted versus actual values, and discuss three or four specific errors. Close with limitations: data size, possible bias, conditions where the model would not hold.

This structure takes less space than pages of screenshots, and it answers the examiner's questions before they are asked. It also makes it easy to turn the chapter into a short paper later.

What data and ethics questions should you be ready for?

• Who collected the data, and did you have the right to use it?
• Does the dataset represent the people the system will be used on?
• Could the model treat any group unfairly, and how did you check?
• How is personal data stored, anonymised or deleted in your demo?
• What should a user do when the model is wrong?

Which deployment options suit a student demo?

Keep the demo simple and local unless your topic is about deployment. A Streamlit or Flask app running on your laptop, with the trained model saved to a file, is the most reliable choice for a viva room with uncertain internet. If you want a public link, a free tier on Hugging Face Spaces or Render works for small models. Whatever you choose, rehearse the demo on the laptop you will carry, with sample inputs ready, and keep a short screen recording as a backup.

Frequently asked questions