The best FYP ideas for AI students solve one clear problem with data you can actually get, beat a simple baseline, and run inside a working app rather than a notebook. Below are 20 AI project ideas for BSAI, BSDS and BSCS students, each with a data source, a tech stack and a difficulty level.
Want ideas across every field, not only AI? See our 70 FYP ideas for CS students or try the FYP idea generator.
Jump to: Pick by compute · NLP and Urdu · Computer vision · Generative AI and agents · Prediction and analytics · What scores well · FAQ
Pick an AI FYP by the compute you have
The most common reason AI projects fail is not the idea. It is choosing a model the group cannot train. Match the idea to the hardware you really have:
| Compute you have | What fits | Examples from this list |
|---|---|---|
| Laptop CPU only | Classical ML, small models, TensorFlow Lite, RAG on top of a hosted LLM API | Used car price estimator, mobile wallet fraud detector, law search assistant |
| Free cloud GPU (Kaggle notebooks, Google Colab when available) | Fine-tuning small CNNs and BERT-size models, training YOLO on a few thousand images | Crop disease detector, pothole detection, exam difficulty estimator |
| University GPU or paid cloud credits | Speech models, handwriting OCR, fine-tuning 1–3B language models | Urdu voice note transcriber, Urdu handwriting OCR, small language model for FAQs |
NLP and Urdu language FYP ideas
Urdu and Roman Urdu are still under-served by AI tools, so a working Urdu project is both useful and easy to defend as original.
| # | Idea | What it does | Data | Tech stack | Difficulty |
|---|---|---|---|---|---|
| 1 | Urdu voice note transcriber | Turns Urdu voice notes, like the ones customers send on WhatsApp, into text and tags the type of request | Mozilla Common Voice (Urdu) plus your own recordings | Python, Whisper, FastAPI | Hard |
| 2 | Roman Urdu to Urdu script converter | Converts Roman Urdu typed on phones into Urdu script, handling the many spellings of the same word | Public Roman Urdu–Urdu word pairs plus pairs you label | Python, PyTorch (seq2seq or a small transformer) | Medium |
| 3 | Pakistani law search assistant | Answers questions about public laws by quoting the exact section, and says so when it cannot find an answer | Laws published on the Pakistan Code website | Python, embeddings, RAG, React | Medium |
| 4 | Urdu handwriting OCR | Reads handwritten Urdu (Nastaliq) pages into editable text, a well-known hard problem | Public Urdu handwriting datasets plus pages you collect and label | Python, PyTorch, CNN with a sequence model | Hard |
| 5 | Exam question difficulty estimator | Predicts how hard a new exam question will be from past questions and student results | Anonymised past papers and results, with your department’s permission | Python, scikit-learn, BERT | Medium |
Related: Building AI products for Urdu and Roman Urdu
Computer vision FYP ideas
Vision projects demo well. Keep the model small enough to run on a phone or a laptop on defence day.
| # | Idea | What it does | Data | Tech stack | Difficulty |
|---|---|---|---|---|---|
| 6 | Offline crop disease detector | Identifies leaf diseases from a phone photo without internet, for farmers in low-signal areas | Public leaf-disease datasets on Kaggle plus your own field photos | TensorFlow Lite, Flutter | Medium |
| 7 | Pothole detection from dashcam video | Finds potholes in road video and pins each one on a map for the local authority | Public pothole datasets plus your own dashcam footage | YOLO, OpenCV, maps API | Medium |
| 8 | Form field extractor for scanned documents | Finds and reads fields such as name, CNIC number and date on scanned admission or bank forms | Blank and filled sample forms you create yourself | Python, OpenCV, Tesseract or a layout model | Medium |
| 9 | Livestock weight estimator | Estimates an animal’s weight from photos, useful for fair pricing at cattle markets | Photos and scale weights you collect at a farm or mandi | PyTorch, OpenCV, Flutter | Hard |
| 10 | Fruit ripeness grader | Grades mangoes or bananas by ripeness from a photo to help sorting before sale | Photos you collect plus public fruit datasets | PyTorch, OpenCV | Easy |
Related: Computer vision internship · Deep learning internship
Generative AI and agent FYP ideas
LLM projects are popular, so examiners look harder at your own contribution. Build the retrieval, the tools or the evaluation yourself.
| # | Idea | What it does | Data | Tech stack | Difficulty |
|---|---|---|---|---|---|
| 11 | Admission counselling agent | Compares programmes across universities, checks eligibility and runs a merit calculator as a tool, citing the prospectus for each answer | Public university prospectuses and merit formulas | Python, LLM API with tool calling, RAG | Medium |
| 12 | Programming assignment autograder with hints | Runs student code against test cases and gives hints instead of the answer | Your course’s past assignments and test cases | Python, Docker sandbox, LLM API | Medium |
| 13 | Small language model for business FAQs on a CPU | Fine-tunes a small open model on one business’s FAQs so it answers without a paid API | FAQ pairs from a business, with its permission | Python, LoRA, a 1–3B open-weight model, llama.cpp | Hard |
| 14 | Hallucination checker for RAG answers | Checks every sentence of a chatbot’s answer against the retrieved sources and flags unsupported claims | Public question-answering datasets plus your own chatbot logs | Python, NLI model, embeddings | Medium |
| 15 | Urdu voice agent for clinic appointments | Answers appointment calls in Urdu, checks free slots and books them | Test conversations you record yourself | Whisper, LLM API, text-to-speech, calendar API | Hard |
Related: Multi-agent AI systems explained · LLM vs traditional machine learning
Prediction and analytics FYP ideas
Tabular data projects are the safest choice for groups without a GPU, and explainable results go down well in a viva.
| # | Idea | What it does | Data | Tech stack | Difficulty |
|---|---|---|---|---|---|
| 16 | Mobile wallet fraud detector | Flags suspicious transactions in a mobile-money style system and explains why each one was flagged | PaySim synthetic mobile-money dataset (Kaggle) | Python, XGBoost, SHAP, FastAPI | Medium |
| 17 | Crop yield prediction from satellite images | Predicts wheat or rice yield per district from vegetation indices across the season | Free Sentinel-2 imagery (Copernicus) plus published yield statistics | Python, Google Earth Engine or rasterio, scikit-learn | Hard |
| 18 | Essential item price forecaster | Forecasts weekly prices of flour, sugar, ghee and vegetables city by city | Weekly SPI price data from the Pakistan Bureau of Statistics | Python, Prophet or LSTM, Streamlit | Medium |
| 19 | Used car price estimator | Suggests a fair price for a used car from its make, model, year, mileage and city | Public Pakistani used-car listing datasets on Kaggle | Python, LightGBM, React | Easy |
| 20 | Electricity theft detector | Finds consumers whose usage pattern suggests meter tampering, a real problem for distribution companies | Public smart-meter theft datasets such as SGCC | Python, CNN or XGBoost | Medium |
Related: Machine learning roadmap · Machine learning internship
What makes an AI FYP score well
- Data first. Confirm you can get enough data before you write the proposal. If you collect it yourself, get consent and remove personal details.
- A metric chosen in advance. F1 for imbalanced classes, word error rate for speech, mAP for object detection. Say why it fits your problem.
- A baseline you beat. Compare your model with something simple, such as logistic regression or a rule-based system, so the improvement is visible.
- Error analysis. Show where the model fails and why. Examiners trust honest limitations more than a perfect accuracy number.
- A working app. Wrap the model in a mobile or web app. A live demo is worth more than a notebook full of charts.
When you write it up, our guides on the FYP proposal, FYP documentation and viva questions will help.
Frequently asked questions
What are good FYP ideas for AI students?
Good AI FYP ideas solve one clear problem with data you can get, such as an offline crop disease detector, an Urdu voice note transcriber or a mobile wallet fraud detector. The best ones are evaluated against a baseline and run inside a real app.
Do I need a GPU for an AI final year project?
No. Classical machine learning and small models run on a laptop CPU, free GPUs on Kaggle and Google Colab are enough to fine-tune small vision and language models, and LLM projects can use a hosted API. Pick the idea to match the compute you have.
Can I use ChatGPT or another LLM API in my AI FYP?
Many supervisors accept it if your own contribution is clear, for example the retrieval pipeline, the tools the model can call, the evaluation or a fine-tuned model. Calling an API and showing its output is rarely enough, so agree the scope with your supervisor early.
Where can I find datasets for an AI FYP in Pakistan?
Start with Kaggle, Hugging Face Datasets and the UCI Machine Learning Repository. For local data, the Pakistan Bureau of Statistics publishes price and economic data and the Pakistan Code website publishes laws. You can also collect your own data, with consent.
How do I evaluate my AI FYP?
Pick a metric that matches the task, keep a test set the model never sees during training, compare against a simple baseline and include an error analysis. If the model runs on a phone, report its speed on a real device too.
Turn the idea into a finished project
Most AI groups lose time on data cleaning and deployment, not on the model. How Ezitech can help:
- Build it yourself with mentors: an AI internship at Ezitech puts you on hands-on projects with Python, machine learning and LLMs before your defence.
Planning a career in AI after your degree? Read how to become an AI engineer.
FYP ideas by degree: Software engineering · AI · Cyber security · IT · Computer science (70 ideas)
More FYP help: FYP idea generator · FYP proposal format · FYP documentation · FYP viva questions