A final-year project does two jobs. It has to satisfy an examiner in a viva, and — far more useful — it is the project you will be asked about in every interview for the next two years. Most students pick for the first job and regret it in the second: an image classifier trained on a famous dataset impresses nobody, because every interviewer has seen a hundred of them. This guide gives you twelve projects that do both jobs, explains how to choose between them, and shows how to scope, evaluate and write one up so that it stands up to questioning.
How to choose: four questions before you commit
Before you fall in love with an idea, put it through four questions. A project that fails any of them will cost you a semester and give you little to talk about.
- Can you get the data? Public documents, your own college's notices, a product's help pages, open datasets. If the data is "we will collect it later", pick something else.
- Can you finish it in the time you have? A working small system beats an ambitious half-built one every time. The hour estimates below are for one person who already knows Python.
- Can you show it works? The strongest projects have an evaluation — a set of test questions with known answers, or a measured accuracy — not just a demo that looked right once.
- Can you explain a decision you made? Interviewers ask "why did you choose X over Y?". Pick a project where you will have to make real choices, such as how to split documents or when the system should refuse to answer.
Beginner: finish in a few weekends
These are the right choice if your programming is still developing, or if your project carries only a few credits. Each is small enough to finish properly, which matters more than ambition.
- PDF Question-Answering Agent (8–14 hours). Ask questions about a set of PDFs and get answers with citations to the page. The classic retrieval-augmented generation (RAG) project, and a strong one if you add an evaluation and make it refuse when the answer is not in the documents.
- Meeting Transcript to Tracked Actions (6–10 hours). Turn a meeting transcript into a list of decisions, owners and deadlines in a fixed structure. Teaches structured output — getting a language model to return data your code can trust.
- Resume and Job Description Matcher (8–12 hours). Compare a resume with a job description and explain the gaps. Easy to demonstrate, and every interviewer immediately understands the problem.
- LLM Cost and Latency Observability (6–12 hours). A wrapper that records what every model call costs and how long it takes, with a simple dashboard. Unusual, practical, and a good answer to "how would you run this in production?".
Intermediate: a full semester project
The best fit for most final-year projects. Each involves a real design decision you will be able to defend.
- Tool-Using Research Agent (12–20 hours). An agent that searches the web, reads sources and writes an answer with citations. Shows you understand how agents decide which tool to call — the most asked-about topic in AI interviews right now.
- An MCP Server Over Your Own Documentation (6–12 hours). Expose your college's or a project's documents to AI assistants through the Model Context Protocol. Small, current and rare among student projects.
- Support Ticket Triage and Routing (10–16 hours). Classify incoming support messages, extract the key details and route them to the right team, with a confidence threshold for sending uncertain ones to a person.
- A RAG Evaluation Harness (8–14 hours). A test suite that scores any question-answering system on accuracy and faithfulness to its sources. Pairs well with the PDF agent if you want two connected chapters.
- Invoice Extraction with a Confidence Gate (10–16 hours). Read invoices into structured data and flag the ones the system is unsure about. A real business problem with a measurable result.
- Semantic Search Over a Codebase (10–16 hours). Search a code repository by meaning rather than keywords. A good choice if you are aiming at software engineering roles.
Advanced: for strong programmers and team projects
Pick one of these only if you are comfortable debugging something that fails in non-obvious ways. Done well, they are the most impressive projects on this list.
- Read-Only SQL Agent (14–22 hours). Answer questions about a database in plain English, with safety built in so it can never change data. Security and correctness give you plenty to discuss.
- Multi-Agent Content Pipeline with a Critic (16–26 hours). Several agents research, write and review a piece of content, with a critic that sends work back. Good for a team of two or three, each owning an agent.
What examiners and interviewers actually look for
Both audiences reward the same things, and neither cares much about how fancy the model is.
- A clearly stated problem, with who has it and why the obvious approach is not enough.
- An architecture diagram you can draw from memory and explain box by box.
- An evaluation: a test set, a number, and an honest comparison — before and after a change, or against a simple baseline.
- Failure cases you found and what you did about them. "It sometimes made up answers, so I added a refusal when no source matched" is worth more than any accuracy figure.
- Code someone else could run, with a README that says how.
Turning the project into a report
Most project reports follow a similar chapter structure. Plan for it from the start, so the writing is a record of work you have already done rather than a reconstruction at the end.
- Introduction and problem statement: the problem, who has it, and what you set out to build.
- Literature and existing approaches: what others do, and the limitation you address.
- System design: the architecture diagram and the key decisions, each with the alternative you rejected.
- Implementation: the components and the tools used, briefly — code belongs in the appendix or repository.
- Evaluation and results: the test set, the metrics, the numbers, and what surprised you.
- Limitations and future work: what it cannot do yet. Examiners trust students who know the limits of their own work.
Questions to prepare for the viva and interviews
Whichever project you choose, prepare a short spoken answer to each of these. They come up in vivas and job interviews alike.
- Why did you choose this approach over a simpler one?
- How do you know it works? What would make the evaluation more convincing?
- What was the hardest bug, and how did you find it?
- What happens when the input is unusual, empty or malicious?
- If you had another month, what would you change first?
Where to find the full specifications
Every project above is specified in full in the AgenticX Project Vault — the problem, architecture, build steps, evaluation, failure modes and an interview explanation for each. Three of them — PDF Question-Answering Agent, Tool-Using Research Agent, and An MCP Server Over Your Own Documentation — are free to read in full on the Projects page, with no sign-up. Start with one of those; if the format helps, the other nine are in the Vault.
If you are still unsure what to build, the free AI Project Idea Generator suggests five projects matched to your level, interests and available hours, each with how you would evaluate it.
Key takeaways
- 1.Choose a project you can get data for, finish in time, evaluate and defend — in that order.
- 2.An evaluation and honest failure cases impress examiners and interviewers more than a bigger model.
- 3.Beginner projects done properly beat advanced projects left half-built.
- 4.Plan the report chapters from the start so writing it records work already done.
- 5.Three of the twelve full specifications are free to read on the Projects page.
Do it now, free
AI Project Idea Generator
Five ideas, each with an architecture, a realistic number of hours, and how you would evaluate it — because the difference between a portfolio project and a tutorial clone is usually evaluation.
No sign-up. The internship is free to join too, with real project briefs and reviewed submissions.