Machine Learning Internship

Classical ML done properly — including the evaluation everyone skips.

A grounding in supervised learning where the emphasis is on the parts that decide whether a model is useful: honest validation, class imbalance, feature leakage and metrics that match the business question rather than the default in the tutorial.

Beginner-friendly4 weeks · remote4 briefs · complete any 2Free to join

What you'll be able to do

Build a defensible classification pipeline end to end
Choose metrics that match the problem, not the tutorial
Spot data leakage before it flatters your results
Communicate a model to a non-technical reader

Who it suits

Students beginning in ML who want to build something they can actually defend in an interview.

What you'll work with

Python
scikit-learn
pandas
NumPy
Matplotlib
XGBoost

The project briefs

Complete any 2 of these 4to finish the programme. Assessment criteria are published here so you know what “good” means before you start.

Credit Scoring Model

~10h estimated

Predict an applicant's creditworthiness from historical financial data.

Engineer features from financial history, then compare logistic regression against a tree ensemble. Note that accuracy is close to meaningless here — the classes are imbalanced and the two error types cost very different amounts.

Assessed against

Precision, recall, F1 and ROC-AUC reported, not accuracy alone
Class imbalance explicitly addressed
A written argument for which error is worse and why
Feature importance explained in plain language

Suggested datasets

UCI German Credit
Give Me Some Credit (Kaggle)

You'll finish with

GitHub repo with notebook
A one-page summary for a non-technical reader

Disease Prediction from Medical Data

~10h estimated

Estimate disease likelihood from structured patient data.

Compare several classifiers on a clinical dataset. In this domain a false negative is far more costly than a false positive — your threshold choice must reflect that, and you must say so.

Assessed against

At least three algorithms compared fairly
Threshold chosen deliberately, with the reasoning written down
Confusion matrix interpreted in clinical terms
A stated limitation section — what this model must never be used for

Suggested datasets

UCI Heart Disease
Pima Indians Diabetes
Breast Cancer Wisconsin

You'll finish with

GitHub repo
Model card covering intended use and limitations

Handwritten Character Recognition

~12h estimated

Classify handwritten digits or characters with a CNN.

Start from a simple baseline before reaching for depth. Report what each architectural change actually bought you — most of them buy less than expected.

Assessed against

A non-neural baseline to compare against
Training curves showing over- or under-fitting
Misclassified examples inspected and discussed
Inference time measured

Suggested datasets

MNIST
EMNIST

You'll finish with

GitHub repo
Error analysis of the 20 worst predictions

Emotion Recognition from Speech

~14h estimated

Classify the emotion in a spoken audio clip.

Extract MFCC features and train a sequence or convolutional model. Split by speaker, not at random — otherwise the model learns voices instead of emotions and your score is fiction.

Assessed against

Speaker-independent train/test split, explicitly justified
Feature extraction explained
Per-class performance, since some emotions are far harder
An honest discussion of what the model confuses

Suggested datasets

RAVDESS
TESS
EMO-DB

You'll finish with

GitHub repo
Per-class results table

Free to apply and free to complete. The completion certificate is optional.

Other domains

AI Agents & Automation

Build autonomous agents that use tools, not chatbots that answer questions.

Generative AI & LLMs

Go past prompt tinkering into retrieval, evaluation and cost control.

Data Science

Turn a messy dataset into a decision somebody can act on.

Ready to Transform Your Enterprise with AI?

Book a free 30-minute consultation with our AI architects. Discover how agentic AI can automate your workflows and multiply your team's output.

No commitment required · Free architecture review · NDA protected

Machine Learning Internship

Classical ML done properly — including the evaluation everyone skips.

Beginner-friendly4 weeks · remote4 briefs · complete any 2Free to join

What you'll be able to do

Build a defensible classification pipeline end to end

Choose metrics that match the problem, not the tutorial

Spot data leakage before it flatters your results

Communicate a model to a non-technical reader

Who it suits

Students beginning in ML who want to build something they can actually defend in an interview.

The project briefs

Complete any 2 of these 4to finish the programme. Assessment criteria are published here so you know what “good” means before you start.

Credit Scoring Model

~10h estimated

Predict an applicant's creditworthiness from historical financial data.

Assessed against

Precision, recall, F1 and ROC-AUC reported, not accuracy alone
Class imbalance explicitly addressed
A written argument for which error is worse and why
Feature importance explained in plain language

Suggested datasets

UCI German Credit
Give Me Some Credit (Kaggle)

You'll finish with

GitHub repo with notebook
A one-page summary for a non-technical reader

Disease Prediction from Medical Data

~10h estimated

Estimate disease likelihood from structured patient data.

Compare several classifiers on a clinical dataset. In this domain a false negative is far more costly than a false positive — your threshold choice must reflect that, and you must say so.

Assessed against

At least three algorithms compared fairly
Threshold chosen deliberately, with the reasoning written down
Confusion matrix interpreted in clinical terms
A stated limitation section — what this model must never be used for

Suggested datasets

UCI Heart Disease
Pima Indians Diabetes
Breast Cancer Wisconsin

You'll finish with

GitHub repo
Model card covering intended use and limitations

Handwritten Character Recognition

~12h estimated

Classify handwritten digits or characters with a CNN.

Start from a simple baseline before reaching for depth. Report what each architectural change actually bought you — most of them buy less than expected.

Assessed against

A non-neural baseline to compare against
Training curves showing over- or under-fitting
Misclassified examples inspected and discussed
Inference time measured

Suggested datasets

MNIST
EMNIST

You'll finish with

GitHub repo
Error analysis of the 20 worst predictions

Emotion Recognition from Speech

~14h estimated

Classify the emotion in a spoken audio clip.

Extract MFCC features and train a sequence or convolutional model. Split by speaker, not at random — otherwise the model learns voices instead of emotions and your score is fiction.

Assessed against

Speaker-independent train/test split, explicitly justified
Feature extraction explained
Per-class performance, since some emotions are far harder
An honest discussion of what the model confuses

Suggested datasets

RAVDESS
TESS
EMO-DB

You'll finish with

GitHub repo
Per-class results table

Free to apply and free to complete. The completion certificate is optional.