Machine learning engineering internships: roles and a project to build

Choose the part of the system you can demonstrate

Read the responsibilities, not only the AI label. A role evaluating recommendations calls for a comparison of model errors. A serving role needs evidence about requests, failures and resource use. A research role may ask for a particular scientific background. Tailor one project to that work rather than attaching every notebook you have.

For a first applied-ML project, start with a decision a person would make from the output. The fictional example below routes support requests for review. Its purpose is to show evaluation reasoning, not to claim that a real model achieved these results.

A 90%-accurate model can miss every urgent request

Suppose 10 of 100 labeled requests are urgent. The baseline marks everything nonurgent. An alternative flags 17 requests: eight urgent and nine ordinary. Work out the mistakes before choosing the model.

• The alternative has lower accuracy but finds 80% of the urgent requests. Accuracy alone would select the baseline that never does the intended task.
• Assume an urgent request missed costs five units and an ordinary request flagged costs one. The baseline costs 10 × 5 = 50; the alternative costs 2 × 5 + 9 = 19. These are invented decision costs, not dollars or employer estimates.
• Change the missed-request cost to half a unit and the comparison reverses: 5 versus 10. The best choice depends on the consequences and review capacity, not the model name.

Turn the example into work you can defend

For your own project, keep a simple baseline, record predictions on a held-out set and produce a table of both kinds of error. When requests belong to the same conversation or customer, check whether a row-by-row split would expose closely related material during training. Define the split before choosing the winning model; repeated tuning on the final test set weakens the comparison.

Package the task, split rule, baseline, error table and one limitation in a short README, with a script that reproduces the calculation. Identify any hosted or pretrained model you used and the work you added around it. A useful interview answer is why a seemingly better headline score would lead to the wrong decision.

Ask the team which result an intern owns: a model, an evaluation, a serving component or a paper. Then ask who reviews the experiment. Those answers tell you whether your project matches the internship.

Before you apply

Eligibility

NVIDIA lists BS, MS, PhD and MBA internships; individual ML roles can require a narrower degree or subject.

Timing

NVIDIA describes year-round internships of at least 12 weeks. Stripe generally schedules internship interviews around local academic calendars.

Location

Check the team location and available data or compute before committing to a project or move.

Pay

Use the rate for your degree level and office. An annualized salary is not the cash paid for a summer.

Sources: NVIDIA university recruiting · Stripe emerging talent

Job titles to search

Applied ML engineering intern

Improve a prediction, ranking or classification used in a product.

ML infrastructure or inference intern

Make training, serving or evaluation cheaper, faster or more dependable.

ML research intern

Investigate a method or research question; read the academic requirements first.

Where to look

Search routes, not confirmed vacancies.

Compare the work and try an example

Use a role comparison or worked exercise to choose your preparation before tailoring an application.

• Internship skills: distinguish required, preferred and taught on the job
• Build an economic consulting work sample that another person can reproduce
• Citadel versus Citadel Securities internships: choose the employer and role