Data Science Portfolio Projects That Get Interviews in Kenya

McTaba runs coding and data bootcamps in Kenya. This guide is written to be useful whether or not you ever join one.

Picture a hiring manager at a Nairobi lender with forty applications open. Thirty-nine say "proficient in Python". One links to a repository that scores loan repayment on a chama's records and runs in a browser tab. She opens that one first. That is the whole idea.

The reason is in the ads. Of 62 Kenyan data job ads we read on 3 November 2026, only six carried a junior or intern title, and the middle lower bound of years asked, among the 52 that stated one, was three (MyJobMag, checked on 3 November 2026). You will not have the years. A project that anyone can open and check stands in for them. We make no claim about what any named recruiter says, only about what the ads ask for.

For the whole path from first lesson to first job, see the data science career guide for Kenya. For general developer portfolios, our developer portfolio guide covers the basics.

Project	Question it answers	What you ship
1. Your M-Pesa statement	Where does the money go, and when does it run out?	A notebook and a monthly chart
2. A chama credit score	Which member is likely to miss the next contribution?	A scoring model with a fairness check
3. A shop's stock forecast	How many units of sugar will sell next week?	A forecast and a reorder alert
4. Mobile money fraud	Which transactions look unlike the rest?	A classifier behind an API
5. A Kenyan public dataset	What does a government or NGO table show that headlines miss?	A dashboard and a short write-up

Want to build these with feedback and a deadline? The Data Science to AI Engineering programme has a capstone on a financial machine learning product. Below is how to build each yourself.

1. Analyse your own M-Pesa statement

Start with your own statement, since it is data you already hold. Request it from Safaricom, load it, parse the dates and amounts, and group by counterparty and month. Follow our guide to analysing an M-Pesa statement in Python. Never publish a real statement. Use your own, with names and numbers masked, or a synthetic one.

2. Score a chama

Chama contribution books are a ready-made credit problem: who paid, who paid late, who borrowed. Build a simple scoring model, then test it for unfair patterns across members. The steps are in our chama credit scoring guide. Use made-up members. Real chama records are other people's personal data.

3. Forecast a shop's stock

A till or a stock book gives dates, items and quantities. Forecast weekly sales for a few fast items, then add a rule that flags when stock will run out. See forecasting shop stock from till data.

4. Detect mobile money fraud

Fraud is rare, so accuracy lies: if one transaction in a hundred is fraud, a model that says "all fine" is right ninety-nine times in a hundred and useless. Practise precision, recall and thresholds, then wrap the model in an API. Our walkthrough is mobile money fraud detection with machine learning.

5. Work a public dataset

Pick a Kenyan table from the Kenya National Bureau of Statistics (KNBS), the Humanitarian Data Exchange Kenya page (HDX Kenya) or a competition on Zindi. Read each dataset's licence before you publish. Our guide to Kenyan open datasets lists where to look and what to check.

A good project fails if nobody can understand it in two minutes. Give every repository three things.

• A README that answers five questions: what is the question, where did the data come from, what did you do, what did you find, how do I run it. Put the answer to "what did you find" in the first three lines.
• A deployed demo. A Streamlit page, a small FastAPI endpoint or a notebook on Google Colab and Kaggle all work. A reader who cannot run your code in one click will not run it.
• A one-page write-up. A blog post or a README section with one chart, the main number, the limits, and what you would do next.

A starting README skeleton:

# Chama repayment scoring

**Finding:** Members with two late payments in the first quarter were the likeliest to miss the next one.

## Question
Can late contributions predict a missed one?

## Data
Synthetic contribution book, 40 members, 12 months. Built by me. No real names.

## Method
Logistic regression on lateness counts. Checked precision and recall.

## Limits
Small, synthetic data. Not a lending decision.

## Run it
pip install -r requirements.txt
streamlit run app.py

Notice the "Limits" section. Stating what your model cannot do is the clearest sign of someone who has worked with real data. For the story side, see our data storytelling guide.

• Using the Titanic or the Iris flowers. Every applicant has them. A Kenyan dataset is a differentiator at no extra cost.
• Publishing real personal data. A real M-Pesa statement or chama book contains other people's phone numbers and names. Mask it or make it up.
• Ten half-finished notebooks. Finish three. A finished project has a README, a demo and a write-up.
• A model with no baseline. Say how well a simple average would do, then show your gain.
• No tests of the number. If your fraud model catches 90 of 100, say how many good transactions it blocked too.

The Data Science to AI Engineering programme starts on 11 January 2027 and costs KES 140,000. Its phases end with a team capstone, deployed and shown at a public demo day. No school can promise a job, and our first cohort has not started. You can build all five projects above without us. If you would rather build them with a cohort, see Data Science to AI Engineering.