Final Year Project Ideas for Computer Science (2026)
A list of project directions that are narrow enough to finish, mapped to the courses you have already taken and the data you can realistically get.
A good final year project is narrow, testable, and built on data or infrastructure you already have access to. The strongest 2026 directions cluster around applied machine learning, campus systems, and security tooling, because all three let you show a measurable result inside two semesters. Pick an idea where you can name the dataset, the hardest technical component, and the user before you write a single line of code.
Most final year projects do not fail because the idea was too simple. They fail because the idea was too wide, the data never arrived, or nobody could tell whether it worked. The list below is organized around that reality rather than around what sounds impressive in a proposal.
Every direction here can be finished by one or two students in two semesters, and each one maps to a course you have probably already taken. Where a real student project on MyUni already exists, it is named so you can see how someone else scoped it.
How to read this list
Each idea below carries three things: data you need to obtain, the piece that will actually be hard, and how you prove it works. If you cannot fill in all three for an idea, it is not a project yet. It is a topic.
Before committing, spend two weeks building the ugliest possible version of the hardest component. Not the login page. The hard part. If the hard part works even badly in two weeks, the project is feasible. If it does not, you have saved yourself a semester.
Applied machine learning
Machine learning projects are popular because they look impressive, and they fail more often than any other category because students pick the model before they pick the data. Reverse that order.
Research paper recommendation
Build a system that recommends papers to a student based on their reading history. The data problem is real but solvable: use a public corpus such as arXiv or Semantic Scholar, and treat the recommendation as a ranking problem rather than a classification one. The hard part is building a usable representation of a paper, usually title plus abstract embeddings. You prove it works with precision at k against a held out set of citations.
A project with this scope already exists in the MyUni showcase.
Urdu or regional language speech recognition
Speech recognition for under resourced languages is genuinely useful and genuinely hard, which makes it a strong academic project. The catch is data. You will spend more time collecting and cleaning audio than training. Scope it to a narrow domain such as digits, commands, or a fixed set of phrases, and you can produce a working model with a clear word error rate. Attempting open domain transcription will not finish.
Crop disease detection from leaf images
Public datasets of crop leaf images are widely available, which removes the hardest part of most ML projects. The technical work is transfer learning on a pretrained vision model plus a clean evaluation across disease classes. Where you can add original value is in deployment: a model that runs offline on a phone in a field is worth more than a notebook with high accuracy. A project like this exists in the showcase.
Reporting 98 percent accuracy on a dataset where 98 percent of samples belong to one class is not a result. Always report alongside a baseline. If a trivial model gets the same score, you have not shown anything.
Campus and institutional systems
These projects are underrated. You have direct access to users, you can run a real pilot, and the problem is concrete enough that nobody has to imagine the use case. Several of the most complete projects in the MyUni showcase fall into this group.
Smart attendance using face recognition
The technical component is a face recognition pipeline. The interesting part is not the model, it is the constraints: lighting in a real lecture hall, students who look similar, and the privacy question of storing biometric data. Handle those explicitly and you have a project with genuine depth. Prove it works with measured false accept and false reject rates on your own captured set.
Campus event management
Unfashionable and finishable, which is a good combination. Event creation, RSVPs, reminders, and an organizer dashboard. Add one non trivial piece to lift it above a tutorial: seat allocation, conflict detection across rooms, or an attendance check in flow. Without that piece it will read as a coursework assignment.
Peer to peer textbook exchange
A marketplace with a trust problem, which is what makes it interesting. Listings, search, and messaging are routine. The hard parts are matching supply to demand across campuses and handling disputes when a book never arrives. A reputation system with a clear model of how trust accumulates is a legitimate research contribution.
Smart parking
Detection of free spaces from camera input or simulated sensor data, plus a booking flow. If you cannot get real camera access, build a simulation and say so plainly. A simulation with honest assumptions is defensible. Pretending simulated data is real is not.
Security and systems
Security projects attract students who want something that looks serious, and then stall because the scope was never bounded. Bound it to one attack or one defence.
AI plagiarism detection
Detecting machine generated text is a moving target, and that is exactly what makes it a good project. Treat it as a classification problem with a clearly stated limitation: your detector will have a false positive rate, and you should measure it. Compare a statistical approach against a transformer based one and report where each fails. Do not claim perfect detection. Claim a measured error rate.
Blockchain based e-voting
Heavily proposed and rarely finished well. The reason is that voting has hard requirements unrelated to blockchain: anonymity, verifiability, and coercion resistance. Be honest about which of those your system achieves. A prototype that demonstrates verifiable tallying on a small scale, with a written analysis of what it does not solve, is a better submission than a system claiming properties it cannot deliver.
Online exam proctoring
Detection of unusual behaviour during an online exam, such as multiple faces or gaze leaving the screen. This is a privacy sensitive area, so treat the ethics as part of the deliverable rather than a paragraph at the end. Technically it is a computer vision pipeline with a real time constraint, which gives you a performance number to defend.
Narrowing it down
Cut your list to three candidates. For each, answer these five questions on one page.
| Question | Why it matters |
|---|---|
| Do I have the data today? | Data that arrives later never arrives. |
| What is the hardest component? | If you cannot name it, you have not understood the project. |
| How do I show it works? | A number beats a feature list in every viva. |
| Who uses it? | Even five real users changes how you build. |
| Would a supervisor find it interesting? | Their engagement is worth more than your initial enthusiasm. |
The MyUni project showcase holds real student work including a smart attendance tracker, a crop disease detector, a paper recommendation engine, and an exam proctoring system. Reading how other students scoped the same ideas will tell you more than any list, including this one.
Once you have three candidates, the next step is the one most students skip: choosing between them with a clear head rather than by enthusiasm. That decision process is worth its own guide.
Frequently asked questions
How many ideas should I shortlist before choosing an FYP topic?+
Shortlist three and no more. Three is enough to compare feasibility, and shortlisting ten usually means you have not defined what you actually want to build. For each one, write down the data you need, the hardest technical piece, and who would use it.
Is an AI or machine learning FYP harder than a web application?+
Not automatically, but it fails differently. Web applications usually fail on scope. Machine learning projects usually fail on data. If you cannot name a dataset you can legally access before you start, pick a different idea. A small model on data you actually have beats a large model on data you hope to collect.
Can I reuse an open source project for my final year project?+
You can build on open source libraries, and you should. You cannot submit someone else's finished application as your own. The line is whether you can explain every design decision and defend the parts you wrote. If you cannot explain why a component exists, it does not belong in your submission.
What makes an FYP idea easy to defend in the viva?+
A measurable result. Projects that can show a before and after number are far easier to defend than projects that only demonstrate features. Accuracy on a held out test set, response time under load, or a user study result all give you something concrete to stand behind.
Start your project on MyUni
Browse notes from Operating Systems, Machine Learning, Computer Vision and more, or publish your own work to the project showcase so the next batch can learn from it.