15 Machine Learning Project Ideas for BSc CSIT and BE Computer Students
15 machine learning project ideas for BSc CSIT and BE Computer students with data sources, algorithms, difficulty and a twist to stand out.
On this page
Machine learning projects are popular because they give you a clear result to show: an accuracy number, a confusion matrix and a working demo. They are also easy to do badly. A notebook that trains a model on a famous dataset and prints "95% accuracy" does not tell the examiner what you understood.
This list gives you 15 ideas with the data to use, the algorithm to explain, a difficulty level and a twist that makes the project yours. For data, we name only well-known public datasets. Wherever the data would be about Nepal, the honest approach is to collect your own and say how you did it. Difficulty assumes a team of two to four students.
How to read the table
- Data: either a well-known public dataset or "collect your own".
- Algorithm: the main method you should be able to explain on paper.
- Level: Easy (a few weeks), Medium (a full semester), Hard (needs more data work or deep learning).
The 15 ideas
Scroll sideways to see all columns.
| # | Idea | Data | Algorithm | Level | Make it stand out |
|---|---|---|---|---|---|
| 1 | Nepali news category classifier | Collect your own headlines from public news sites, label them by category | TF-IDF with Naive Bayes or SVM | Medium | Build a Nepali stop-word list and compare it with no stop-word removal |
| 2 | Spam SMS and message detector | UCI SMS Spam Collection (public) plus a few Nepali and Roman Nepali messages you write yourself | Naive Bayes vs logistic regression | Easy | Handle Roman Nepali spelling variations |
| 3 | Plant leaf disease detector | PlantVillage (public) and your own leaf photos from a local field with permission | CNN with transfer learning (MobileNet) | Hard | Test on your own photos and report how accuracy drops |
| 4 | Handwritten digit recognizer | MNIST (public) | CNN, compared with k-NN | Easy | Draw digits on a web canvas and predict live |
| 5 | Devanagari character recognizer | Collect your own handwriting samples (public Devanagari datasets also exist: check the source and licence) | CNN | Hard | Show which similar-looking letters are confused |
| 6 | Student result predictor | Anonymised or self-collected marks and attendance data, with consent | Decision tree or random forest | Easy | Show feature importance and discuss fairness |
| 7 | Sentiment analyser for movie reviews | IMDB movie reviews (public) for training; collect your own Nepali or Roman Nepali movie reviews to test transfer across languages | Logistic regression, then a small neural network | Medium | Show how the model fails on sarcasm and mixed language |
| 8 | Heart disease risk predictor | UCI Heart Disease (public) | Logistic regression, random forest | Easy | Use sample records for a classroom demo and compare precision and recall. Keep it separate from medical decisions |
| 9 | Movie or book recommender | MovieLens (public) | Collaborative filtering (user-based and item-based) | Medium | Add a cold-start fallback using content features |
| 10 | Traffic vehicle counter | Record your own short road videos (a safe location, no private information) | Object detection with a pretrained YOLO-type model, then simple tracking | Hard | Count by vehicle type and compare with a manual count |
| 11 | Air quality level predictor | Collect your own readings with a low-cost sensor and Arduino or ESP32 | Time-series regression, such as linear regression with lag features | Medium | Combine hardware data collection with a web dashboard |
| 12 | Crop recommendation helper | Collect your own soil, rain and crop records with the help of a local agriculture office or a farmer; a public crop dataset also exists on Kaggle | Decision tree or k-NN | Medium | Explain why data from Nepal conditions could differ from foreign data |
| 13 | Face mask or helmet detection | Collect your own photos with consent, or use a public face dataset and read its licence | Object detection with transfer learning | Hard | Run the model live from a webcam and report frames per second |
| 14 | Customer or student segmentation | UCI datasets, or your own anonymised survey | K-means clustering | Easy | Choose k with the elbow method and silhouette score |
| 15 | Fake news or misleading headline detector | Collect your own labelled examples of Nepali and English headlines from fact-checking sources | TF-IDF with SVM, then compare with a pretrained transformer | Hard | Report where the model is unsure instead of forcing a label |
Choosing between them
- Short on time? Pick 2, 4, 6, 8 or 14. They are clear and can be completed with free tools in a few weeks.
- Want a Nepal angle? Ideas 1, 5, 11, 12 and 15 become far more convincing when you collect data yourself and describe the process honestly.
- Want hardware too? Idea 11 combines sensors with a model. Idea 10 needs a camera and a laptop or Raspberry Pi.
- Want a real-time demo? Ideas 4, 10 and 13 make good live demos.
A project structure that examiners like
- Problem statement in one sentence, with who benefits.
- Dataset: source, size, classes, how it was cleaned. For your own data, how you collected and labelled it.
- Train, validation and test split done before any experiments.
- Baseline model (for example majority class, or logistic regression).
- Main model and why you chose it.
- Metrics: accuracy alone is not enough when classes are imbalanced, so show precision, recall, F1 and a confusion matrix.
- Error analysis: show five examples the model got wrong, and say why.
- Deployment: a simple Flask, FastAPI or Streamlit app, or a Flutter app using a TensorFlow Lite model.
- Limits and ethics: bias, privacy, and what the model should not be used for.
Viva tips
- Be ready to explain your main algorithm on paper with a tiny example, for instance five training points for k-NN or a three-word document for TF-IDF.
- Know the difference between training accuracy and test accuracy, and why a large gap can indicate overfitting or a shift in the data.
- Explain how you stopped data leakage, for example making sure the same leaf photo or message is not in both train and test sets.
- Say how many samples you have per class. Examiners often ask.
- Do not call something your own result if it came from a tutorial. Say what you changed.
- If you used a public dataset, cite its source and licence in your report.
- Have a backup demo: a short screen recording in case the internet or laptop fails.
- Do not invent rules about formats or marks. Ask your department or supervisor what report format they expect.
Tools that make the work faster
Scroll sideways to see all columns.
| Task | Free tool |
|---|---|
| Notebook and training | Google Colab, Jupyter |
| Classical ML | scikit-learn, pandas |
| Deep learning | TensorFlow or PyTorch |
| Quick web demo | Streamlit, Flask, FastAPI |
| Data labelling | A spreadsheet, or an open-source labelling tool |
| Mobile deployment | TensorFlow Lite with Flutter or Android |
A warning on datasets
Never present data as "Nepali" if you did not collect it or cannot name its source. If a public dataset has no Nepali content, say so plainly and show how you tested on a small sample of your own. Honest limits score better than inflated claims.
FAQ
Do I need deep learning for a good ML project?
No. A well-explained random forest with a clean evaluation is better than a neural network you cannot explain.
How much data do I need?
There is no fixed number. For classical models, a few thousand labelled rows can be enough, while image models usually need more or need transfer learning. Always report your sample count and test on data the model has never seen.
Can I use a pretrained model?
Yes, and it is often the right choice, as long as you say what was pretrained, what you fine-tuned and how you measured the result on your own data.
Is Python the only choice?
It is the most common because of the libraries, but R or Java can work. Use what your team can explain.
How can I connect ML with electronics?
Collect sensor data with an Arduino or ESP32, train a model on a computer and display predictions in a web dashboard. Our ESP32 IoT ideas article lists sensor projects that could feed an ML model.