40 AI & Machine Learning Final Year Project Ideas for 2026 (With Datasets and Tech Stack)

Short answer: the best AI final year projects solve a real, local problem, use data you can actually get, and end in a working demo, not just a notebook. Pick an idea below that matches your skills, check the dataset exists before you commit, and plan for a deployed app (web, mobile or an edge device) by the final evaluation.

Each idea lists a dataset or data plan, a typical stack and a difficulty level: ★ suits a team new to deep learning, ★★ needs solid deep-learning basics, and ★★★ is research-level or hardware-heavy.

How to choose your FYP in one evening

1. Start from a problem someone has, not a model you want to try. "Farmers lose wheat to rust" beats "I want to use a Vision Transformer".
2. Confirm the data first. If you can't download or collect a few thousand labelled examples, pick another idea or plan data collection into semester one.
3. Scope one strong feature. One accurate, well-evaluated model inside a polished app beats five half-working ones.
4. Plan the demo. Examiners remember a live demo on a phone or camera more than any accuracy table.
5. Agree it with your supervisor early, and follow your university's rules on outside help. The work you submit must be your own.

Computer vision

1. Plant disease detection app: photograph a leaf, get the disease and a treatment tip. Data: PlantVillage plus your own field photos. Stack: PyTorch or TensorFlow, MobileNet/EfficientNet, a Flutter app. ★
2. Pakistani number-plate recognition (ANPR) from CCTV. Data: collect and label your own frames. Stack: YOLO (Ultralytics), EasyOCR or PaddleOCR, OpenCV. ★★
3. Helmet and seat-belt violation detection for traffic cameras. Data: your own footage plus public helmet datasets on Kaggle and Roboflow Universe. Stack: YOLO, ByteTrack. ★★
4. Classroom attendance with face recognition and a liveness check. Data: consenting classmates; public face datasets for testing. Stack: ArcFace-style embeddings, a vector index, FastAPI. ★★
5. Retail shelf monitoring that spots empty shelf space and misplaced products. Data: SKU-110K plus store photos. Stack: YOLO and a web dashboard. ★★
6. Pakistan Sign Language to text, letters or common words. Data: published PSL datasets or your own recordings. Stack: MediaPipe hand landmarks with an LSTM or Transformer. ★★
7. Pothole and road damage mapping from dashcam video. Data: the Road Damage Dataset (RDD) plus your own videos. Stack: YOLO, GPS logging, a Leaflet map. ★★
8. Fabric defect detection for textile mills. Data: public fabric-defect datasets plus factory samples. Stack: anomaly detection (autoencoders, PatchCore). ★★★

Urdu and language AI

1. Roman Urdu sentiment analysis for product reviews or social posts. Data: public Roman Urdu datasets (UCI, Kaggle) plus reviews you label. Stack: a scikit-learn baseline, then fine-tuned XLM-RoBERTa. ★
2. Urdu fake-news detection. Data: published Urdu fake-news benchmark datasets. Stack: multilingual BERT fine-tuning, with SHAP or LIME explanations. ★★
3. Urdu speech-to-text for voice notes, with a short summary. Data: Mozilla Common Voice (Urdu). Stack: fine-tuned Whisper. ★★★
4. Urdu/English support chatbot for a real local business. Data: its FAQs and past chats, with permission. Stack: an LLM with retrieval (LangChain or LlamaIndex) and a vector database. ★★
5. Contract explainer that explains a rental agreement in plain Urdu or English, quoting the clauses it relies on. Data: public sample contracts. Stack: an LLM with retrieval and citations. ★★
6. Exam question generator from textbook chapters, tagged by difficulty. Data: your own course notes. Stack: an LLM and a small evaluation set. ★
7. Handwritten Urdu text recognition (OCR). Data: published Urdu handwriting datasets or samples you collect. Stack: CRNN or TrOCR-style models. ★★★

LLMs, RAG and AI agents

1. University handbook assistant that answers questions about rules, fees and deadlines with citations. Data: your university's public documents. Stack: hybrid search (BM25 plus embeddings), a reranker and an evaluation set. ★★
2. Job-application agent that tailors a CV to a job post, drafts a cover letter and tracks applications. Stack: an LLM with structured outputs and a React app. ★★
3. Multi-agent research assistant: one agent searches, one summarises, one fact-checks. Stack: LangGraph or CrewAI and a web search API. ★★★
4. Meeting minutes generator: record, transcribe, summarise and extract action items. Stack: Whisper, speaker diarisation (pyannote) and an LLM. ★★
5. Text-to-SQL for business data: ask a sales database questions in plain English. Data: sample databases such as Northwind or Chinook. Stack: an LLM, SQL validation and guardrails. ★★
6. AI code reviewer for programming assignments, covering style and likely bugs. Stack: an LLM plus static analysis (pylint, ESLint). ★★
7. WhatsApp order-taking bot for a small shop, from catalogue to order confirmation. Stack: the WhatsApp Business Platform test number, an LLM and a database. ★★

Healthcare

1. Chest X-ray abnormality detection with Grad-CAM heatmaps. Data: NIH ChestX-ray14 or CheXpert. ★★
2. Bone fracture detection from X-rays. Data: MURA (Stanford). Stack: a CNN or ViT with Grad-CAM. ★★ We built a research system like this; see the bone fracture detection case study.
3. Skin lesion classification with a phone camera app and a clear "not a diagnosis" notice. Data: HAM10000 or ISIC. ★★
4. Diabetic retinopathy grading from retina images. Data: APTOS 2019 (Kaggle). ★★
5. Brain tumour segmentation on MRI. Data: BraTS. Stack: U-Net with MONAI. ★★★
6. Prescription reader: OCR on a prescription photo, then a check against a drug-interaction database. ★★

Agriculture, climate and cities

1. Lahore smog and AQI forecasting 24 to 72 hours ahead. Data: OpenAQ plus weather data (Open-Meteo or NASA POWER). Stack: gradient boosting or an LSTM, with a dashboard. ★★
2. Crop yield prediction by district. Data: FAOSTAT, weather data and Sentinel-2 vegetation indices (NDVI). ★★
3. Smart irrigation that predicts when to water from soil sensors. Stack: ESP32 sensors, MQTT and a small model. ★★
4. Livestock weight estimation from photos. Data: collect your own photos with measured weights. ★★★
5. Traffic density estimation and signal timing from junction video. Stack: YOLO with tracking, and simulation in SUMO. ★★★

Business and finance

1. Credit-card fraud detection on imbalanced data. Data: the Kaggle credit-card fraud dataset. Stack: XGBoost, resampling and cost-sensitive evaluation. ★
2. Customer churn prediction. Data: the IBM Telco churn dataset. ★
3. Retail sales forecasting with Ramadan, Eid and holidays as features. Data: public store-sales datasets plus your own calendar features. ★★
4. Resume screening and job matching with fairness checks. Data: public resume datasets and job posts. Stack: embeddings, ranking and fairness metrics. ★★
5. Invoice and receipt data extraction into Excel or accounting software. Data: SROIE and CORD. Stack: Donut, LayoutLM or a vision LLM. ★★

Edge AI and robotics

1. Driver drowsiness detection on a Raspberry Pi or NVIDIA Jetson, with an alarm. Stack: MediaPipe facial landmarks, eye-aspect ratio, ONNX or TensorRT. ★★
2. Campus delivery robot with line following and obstacle detection. Stack: ROS 2, a camera, a small YOLO model, Jetson or Raspberry Pi. ★★★

Make your FYP stand out

• Deploy it. Put the model behind an API (FastAPI) and a real interface: a Flutter or React app, a WhatsApp bot, or a camera device. A Hugging Face Space or Streamlit demo is the minimum.
• Evaluate honestly. Report precision, recall and a confusion matrix on a held-out test set, and show where the model fails.
• Use local data. Even a few hundred images or sentences collected in Pakistan make the project original.
• Write a clean README with setup steps, screenshots and results. Recruiters read GitHub.
• Measure speed and cost: latency per prediction, model size, and what it would cost to run.

A two-semester plan

Month · Milestone

1 · Problem statement, literature review, dataset confirmed

2 · Baseline model and evaluation set

3–4 · Better model and a first working app

5–6 · Deployment and testing with real users

7 · Final evaluation, report, poster and demo video

8 · Buffer for fixes and viva preparation

Get help without crossing the line

The Nexariza Learning Hub offers free tutorials, a hands-on AI cohort and final-year project guidance from engineers who build AI for clients: idea selection, architecture and code reviews, and deployment help. Top learners are offered internships on live Nexariza projects; see open roles. For hardware projects, the Nexariza Store has edge AI hardware and IoT kits. Next: how to get an AI or ML internship in Pakistan.

Frequently asked questions

What is a good AI final year project idea for 2026?

A good AI final year project solves a real problem, uses data you can actually get, and ends in a working demo. Strong 2026 choices include a RAG assistant that answers from your university's documents, Urdu speech-to-text for voice notes, plant disease detection on a phone, and Pakistani number-plate recognition.

Which AI final year projects are easiest for beginners?

Projects marked ★ in this list: plant disease detection with PlantVillage, Roman Urdu sentiment analysis, an exam question generator, credit-card fraud detection and customer churn prediction. They have public data and well-documented baselines, which leaves time to build a good app around the model.

Where can I find datasets for an AI final year project?

Kaggle, Hugging Face Datasets, the UCI Machine Learning Repository, Roboflow Universe for images, Mozilla Common Voice for speech, and Google Dataset Search. For originality, add data you collect yourself, with consent wherever people are involved.

How long does an AI final year project take?

Most BS programmes in Pakistan run the final year project over two semesters, roughly eight months. Confirm your data in month one, build a baseline by month two and a working prototype by month four, and keep the last month for testing, the report and the demo.

Can Nexariza AI help with my final year project?

Yes. The Nexariza Learning Hub offers final-year project guidance from engineers who build production AI: idea selection, architecture and code reviews, and deployment help. We mentor; the work you submit stays your own, in line with your university's rules.