Answer in brief
Choose the project by the problem, not by the technology. State the goal in one sentence without naming a model, confirm that you may use the data, measure a simple baseline first, cut the scope until it fits the weeks you have, and decide in advance how the result will be tested. Finish with a short written account of what the system is for, where it fails and what you did not check.
The project is where the course becomes your decision
Lectures and exercises on an applied AI course are chosen for you. The project is not. It is the one part where you decide what to build, with what data, and how to tell whether it worked, and it is usually the piece of work you will show to an employer or a client afterwards.
Most disappointing projects fail before any code is written. The learner picks a technique that sounds impressive, then looks for a problem to attach it to, then discovers in week five that the data does not exist. Stanford’s project page for its CS230 deep learning course notes that application projects, in which you pick a problem that interests you and work out how to apply learning algorithms to it, are by far the most common kind.
Five tests sort a workable AI course project from a wish: a real problem, data you are allowed to use, a baseline to beat, a scope that fits the weeks available and an evaluation you can explain. A sixth belongs at the end, an honest account of the limits.
State the goal in a sentence that names no model
Google’s Introduction to Machine Learning Problem Framing, a short online course of about 45 minutes by its own estimate, begins with an instruction that sounds too simple to matter: state your goal in non-ML terms, as the answer to the question “What am I trying to accomplish?”. Its examples take one line each: detect spam, calculate travel time, identify fraudulent transactions.
Do the same for your idea. “Fine-tune a language model on support tickets” is a method. “Sort incoming support messages into five queues so that urgent ones are answered first” is a goal: it names who benefits and what changes. If you cannot write the second kind of sentence, you have a tool and no project yet.
The same course adds a test that removes many attractive ideas: there is no value in predicting something if you cannot turn the prediction into an action that helps users. Ask what a person would do differently with your system’s output on Monday morning. A tidy score that changes nobody’s work is an exercise, and an exercise is a weaker project than a small, useful tool.
Data you may use, and can describe in writing
Permission comes before quality. Use data you created yourself, a public dataset whose licence covers your purpose, or material from an employer with written consent and personal details removed. Read the licence and the terms on the dataset’s own page; “it was on the internet” is not a licence. If the data describes identifiable people, ask the course or the organisation first how it must be handled.
Then judge whether it can carry the project. Google’s course lists what data for a predictive model should be: abundant, consistent and reliable, trusted, available at the moment of prediction, correct, and representative of the real world. A learner rarely has all six. The honest response is to note which ones are weak and to shrink the claim accordingly.
Count the time as well. Stanford’s CS230 guidance warns that when data needs considerable pre-processing, or you plan to collect it yourself, this is only one part of the expected work but can often take considerable time. A practical rule: if you cannot open a usable sample of the data in the first week, change the idea, not the schedule.
A baseline to beat before anything clever
A baseline is the simplest thing that could do the job: a rule, a keyword list, the current manual process, or a ready-made model used with a one-line instruction. Google’s course calls the non-ML solution the benchmark for deciding whether machine learning is a good fit for the problem, and suggests solving the problem by hand with a heuristic if no such solution exists.
The baseline is what gives your final number a meaning. An illustration with invented figures: a classifier that is right 90 times in 100 sounds strong until you notice that always choosing the most common category is right 88 times. The two-point difference is the result, and it is a modest one. Without the baseline you would have reported 90 and learned nothing.
Build it early. The same course advises starting simple, because most of the work is on the data side, and observes that simple models make a good baseline even if you never launch them. Stanford’s CS230 asks for running code with a baseline model at the milestone stage. Whatever your course requires, a working baseline in the first third of the calendar means you already have something complete to submit.
Cut the scope until it fits the calendar
Count the weeks you really have, then subtract. The last week belongs to the write-up, the one before it to evaluation and error analysis, and at least one will be lost to work, illness or a broken laptop. What remains is the time for data and building, usually far less than the first plan assumed.
Narrow the task instead of thinning the work. One language, one document type, one kind of customer, three categories in place of twelve. Write down the smallest version that would still be a complete project, with data, a baseline, a test and a conclusion, and treat everything beyond it as optional. A finished narrow project teaches more than an unfinished broad one.
Be wary of ideas that depend on things you do not control: a dataset someone has promised, access that needs approval, thousands of examples that must be labelled by hand. Stanford’s milestone guidance looks for evidence that a team has tested any assumptions which might prevent it from completing the project. Test yours in week one, when changing direction is still cheap.
An evaluation you could explain to a colleague
Decide how you will judge the result before you build it. Set aside a portion of the examples as a test set and do not look at it while you adjust the system; a score measured on examples you tuned against tells you nothing about new ones. Choose one main measure that follows from the goal sentence, and be ready to say in plain words what a point of it means.
Google’s course separates two things that learners often merge. Model metrics such as accuracy, precision and recall describe the model. Success metrics describe what you care about, for example whether people take the right action more often. A project that reports the first and discusses the second shows that its author understood the problem, not only the library.
Generated text needs a different kind of test. The course notes that you may need a test set to compare a generative model’s output with known values, for instance checking that summaries resemble human-written ones or that people rate them as good. In practice that means a short rubric of three or four criteria, a fixed set of inputs and, if you can arrange it, a second person to do the scoring.
Limits belong in the write-up, not in a footnote
Research practice offers a ready template. The paper Model Cards for Model Reporting, by Margaret Mitchell and colleagues, proposes that a trained model be accompanied by a short document that discloses the context in which it is intended to be used and how its performance was evaluated. Its suggested sections include intended use, out-of-scope use cases, evaluation data, and caveats and recommendations.
An AI course project can borrow the idea on one page. Say what the system is for and who should use it. Say what it must not be used for. Describe the data and what it lacks. Give the result next to the baseline. Then show three or four real failures and say what they have in common. Examples of failure are more convincing than a sentence admitting that some exist.
Learners often fear that listing weaknesses will count against them. Consider how the page reads to an assessor or an employer: a result can be trusted only when its boundaries are visible. “Works on English invoices from two suppliers; not tested on handwriting” is a finding. Silence about the same gap is an error waiting for someone else to discover.
Check the idea against more than technique
UNESCO’s AI competency framework for students, published in 2024 for school curricula, describes 12 competencies across four dimensions: a human-centred mindset, the ethics of AI, AI techniques and applications, and AI system design. It arranges them on three levels of progression: understand, apply and create. It was not written for adult courses, but it makes a useful mirror.
Hold your idea up to it. A project that only tunes a model exercises one dimension. One that also asks who is affected by a wrong output, whose data is used and how the pieces fit into a working system touches all four. And a project in which you build, test and improve something of your own sits at the create level, where an applied course aims.
Read your own course’s rules before you commit, because rules vary between courses. Stanford’s CS230 page, at the time of writing, says that simply prompting a large language model without additional training does not qualify as a project for that class; another course may welcome exactly that, and Google’s framing course observes that in most cases you will customise a pre-trained generative model, not train your own. Then choose between your candidate ideas with the five tests, and prefer the one whose data you can open today.
Practical checklist
- Write the goal in one sentence that names no model or technique, and say who would act on the result.
- Get permission to use the data in writing, and record its source, licence, date and known gaps.
- Score a simple baseline on held-out examples before building anything more complex.
- Plan backwards from the deadline and define the smallest version that still counts as a complete project.
- End with one page on intended use, out-of-scope use, results beside the baseline and real examples of failure.
Questions and answers
How do I choose between two project ideas for an AI course?
Score each idea against the same five questions: is the problem real, may you use the data, can you build a baseline this week, does the smallest complete version fit the calendar, and can you explain how you will test it. The idea with data already in hand usually deserves to win, even if the other sounds more exciting. Excitement does not survive a month of waiting for access.
Is a project built on a ready-made language model too simple?
It depends on the rules of your course. Some, such as Stanford’s CS230, require training work and say that prompting alone does not qualify. Where the rules allow it, the depth comes from elsewhere: a precise goal, a carefully assembled test set, a comparison with a simpler baseline and an analysis of failures. Those are the parts most easily skipped, and the parts that show judgement.
What counts as a baseline in a machine learning project?
Any simple method that performs the same task and can be scored on the same test examples: a hand-written rule, a keyword match, always predicting the most frequent class, the existing manual process, or an off-the-shelf model with a minimal instruction. Its purpose is comparison. Report its score beside your own, measured in the same way, so that a reader can see how much your work added.
Can I use my employer’s data for a course project?
Only with permission in writing from someone entitled to give it, and normally after personal and confidential details have been removed. Check as well what the course does with submitted work, since reports and code may be read by staff or shared with other students. If permission is slow or uncertain, choose a public dataset with a clear licence. This is general guidance, not legal advice; your organisation’s rules decide.
How large should the dataset be for a course project?
There is no universal figure. Google’s problem framing course says that more relevant examples make a better model, and lists correct and representative data among its other requirements. For a short project, a workable test is whether the held-out examples include every type of case you claim to handle and whether you can check their labels yourself. If either answer is no, narrow the claim before you collect more.