How I Built a Career Pathway Adviser for Mauritian Students

Choosing what to study after the Higher School Certificate (HSC) involves several questions. Which subjects do I perform well in? What interests me? Which career fields should I explore?

For my MSc in Artificial Intelligence with Machine Learning, I built a prototype to bring some of these questions together. The project became a career pathway adviser for Mauritian HSC school-leavers, combining academic information, interests and a small experimental machine learning component.

The aim was to give students a starting point for exploring their options and discussing them with a career adviser.

A student enters their subjects and grades, answers an interest questionnaire and provides some self-reported information. The system combines these inputs to rank seven broad career fields and present the top three suggestions.

Those suggestions are guidance. Course admission requirements still need to be checked separately, and the system has not been validated as a predictor of a student’s future career success.

I used Python for the application logic, Streamlit for the interface and SQLite to store research sessions. Streamlit allowed me to turn the scoring process into an interactive application where a user could enter information and see the resulting recommendations.

The main design decision was how to combine different kinds of evidence. I used a hybrid approach, bringing together transparent scoring rules and an experimental classifier.

Academic information carried the greatest weight. Interests contributed a substantial share, while the experimental component had a smaller role because the evidence supporting it was limited.

The 60/30/10 split was a design choice. Establishing whether it is the most useful balance would require further evaluation with relevant student data.

For the experimental component, I compared Logistic Regression, Support Vector Machine, Gaussian Naive Bayes and Random Forest. These are different methods for learning relationships between inputs and a target category.

One important limitation was the dataset: the experimental classifier used a public demonstration dataset containing 105 records. That is a small evidence base, particularly for a tool intended for Mauritian students. It supported experimentation, but claims about usefulness for that population would require stronger evidence.

This limitation influenced how I positioned machine learning within the application. It contributed one part of the overall score, with the academic and interest components remaining visible in the design.

Testing also involved more than comparing classifiers.

I ran an audit using 5,000 uniformly generated valid inputs. This allowed me to investigate how the combined scoring system behaved across many possible input combinations.

The audit checked whether all seven career fields remained reachable and how much the experimental component affected the leading recommendation. All seven fields were reachable, and adding the experimental component changed the top-ranked field in 8.9% of the generated cases.

That result describes the component’s influence within the simulation. Establishing whether the recommendations help real students make better-informed choices would require a separate study with students and career guidance professionals.

The project helped me separate three questions: does the application work, does the scoring behave as intended, and is the guidance useful to its intended users? Each question needs its own evidence.

It also made the importance of explanation more concrete. A student should be able to understand which parts of their profile contributed to a suggestion and what the result can reasonably tell them. A polished interface alone cannot answer those questions.

A stronger version of the project would need a larger, locally relevant dataset, feedback from students and advisers, and further checks on how recommendations behave across different student profiles. The weighting choices would also need to be examined more closely.

Building this prototype gave me practical experience connecting data, scoring rules, machine learning, storage and an interface. It also gave me a clearer habit to carry into future projects: define what the output means, then design the evaluation around that claim.