AI Project Ideas for Final Year Students

What separates a good AI project from a copied one

Examiners have seen plenty of notebooks that train a classifier on a clean download and report a high accuracy. What they rarely see is a student who can explain why the model fails on a blurred photo, or what happens when a user uploads something unexpected. Treat the model as one component of a working system, and plan for its mistakes from the first week.

Each idea below can be built by a team of three or four over two semesters. Pick one where you can collect at least some of your own test data, because evidence gathered outside the training set is the most convincing thing you can show in a viva.

Computer vision projects

1. Crop leaf disease detection that works in the field — PyTorch, EfficientNet, Flutter
   Models trained on plain-background leaf images often fail on farm photos. Train with heavy augmentation, then test on photos you take yourself and report both results honestly.
2. Pothole mapping from dashcam video — YOLO, OpenCV, Leaflet
   Municipal road surveys are slow and infrequent. Detect potholes frame by frame, remove duplicate sightings using GPS, and plot a severity map for a ward.
3. Safety gear compliance on construction sites — YOLO, ByteTrack, FastAPI
   Supervisors cannot watch every camera. Detect helmets and vests, track each person across frames, and alert only after a sustained violation to limit false alarms.
4. Indian Sign Language to text — MediaPipe, LSTM or Transformer
   Extract hand and body landmarks rather than raw pixels, then classify word-level gestures with a sequence model. Landmarks reduce sensitivity to background and skin tone.
5. Face attendance with liveness checks — FaceNet, OpenCV, Django
   Basic face attendance is fooled by a phone photo. Add a blink or texture-based anti-spoofing step and measure how often printed and on-screen attacks get through.
6. Fall detection for elderly care using pose only — MediaPipe Pose, scikit-learn
   Cameras in homes raise privacy concerns. Keep only skeleton keypoints, discard the video frames, and classify falls against sitting and bending motions.

Language and speech projects

1. Civic complaint router for code-mixed text — Transformers, FastAPI, MySQL
   Citizens write in a blend of Hindi, Marathi and English. Classify each complaint by department and urgency, and compare a multilingual transformer against a simple TF-IDF baseline.
2. Voice form filling for low-literacy users — Whisper, spaCy, text-to-speech
   Government and bank forms exclude many people. Ask questions aloud, transcribe answers, extract fields such as name and date, and read them back for confirmation.
3. Lecture summariser with timestamps — Whisper, sentence embeddings, React
   Students rewatch long recordings to find one concept. Transcribe the lecture, segment it by topic, and link each summary point back to the moment it was said.
4. Resume and job description matcher with a bias audit — Sentence Transformers, Python
   Rank resumes by semantic similarity, then swap names and colleges in test resumes to measure whether the ranking shifts when qualifications stay the same.
5. Claim checking assistant for news — NLI models, Elasticsearch
   Rather than labelling whole articles as fake, isolate individual claims and retrieve supporting or contradicting passages from a trusted archive, leaving the judgement visible to the reader.

Prediction and decision-making projects

1. Early warning system for student dropout — LightGBM, SHAP, Streamlit
   Use attendance and internal marks to flag at-risk students while there is time to act. Explain each flag so a mentor knows what to discuss.
2. Adaptive traffic signals with reinforcement learning — SUMO, Stable-Baselines3
   Fixed timers waste green time at uneven junctions. Train an agent in a traffic simulator and compare average waiting time against a fixed-time plan.
3. Mandi crop price forecasting for farmers — Prophet, XGBoost, Flask
   Farmers decide when to sell with little information. Forecast short-term prices, show a range instead of a single number, and present it in the local language.
4. Machine fault prediction from vibration signals — Python, 1D CNN, FFT
   Unplanned breakdowns are costly. Convert sensor readings into spectrograms or statistical features, classify fault types, and test on a load condition the model has not seen.
5. Personalised exam practice planner — Knowledge tracing, Django
   Model what each learner knows from their answers, then choose the next question to balance revision and new material. Compare it against random question order.

Edge AI and privacy-focused projects

1. Driver drowsiness alert on a Raspberry Pi — TensorFlow Lite, Raspberry Pi
   Cloud inference is unusable on highways with poor signal. Run eye and yawn tracking on the device, and report frame rate as well as accuracy.
2. Automated waste sorting bin — MobileNet, Arduino, TFLite
   Classify items as wet, dry or recyclable with a small camera and move a servo flap. Quantise the model and measure what accuracy you lose.
3. Federated learning across simulated hospitals — Flower, PyTorch
   Patient records often cannot leave a hospital. Train a shared model without pooling data, and study how uneven data across sites hurts performance.
4. Adversarial testing of a traffic sign classifier — PyTorch, Adversarial Robustness Toolbox
   Small stickers can mislead vision models. Generate attacks, measure the accuracy drop, and evaluate whether adversarial training restores robustness.

How to narrow your choice

• Name the user in one sentence. If you cannot, the idea is still a technique looking for a problem.
• Check data before committing. Spend a day finding a usable public dataset and confirming its licence allows academic use.
• Match the idea to your hardware. A free cloud GPU session suits transfer learning, but training large models from scratch is unrealistic.
• Decide how you will demonstrate failure. Showing a confidence threshold and a fallback route looks more mature than a perfect scripted demo.
• Write down one comparison. Baseline versus your method gives your report a results chapter with real content.