We review portfolios every week when interns apply and again when they go on the market. The gap between projects that get callbacks and projects that don't has very little to do with model complexity. It is about whether a busy person can tell, in ninety seconds, that you can do the job.
The ninety-second test
A hiring manager opens your repository and asks three questions: What problem is this? What did you find? Would I trust this person with our data? Your README has to answer all three above the fold. Lead with the question and the result, in plain language, then the approach. Put the notebook link second, not first.
Four project types that work
1. A messy, real dataset you cleaned end to end. Not Titanic, not Iris. Public city data, a scraped listings site, an open API with missing fields and inconsistent dates. Document every cleaning decision. This is 60% of the real job and almost nobody shows it.
2. A business metric with a decision attached. "Which customers should get a retention offer, given the offer costs $12?" A model whose output is a ranked list with an expected-value calculation beats a model with a higher AUC and no decision.
3. An experiment you designed and analysed. A/B test analysis with power calculation, a clear null, and an honest write-up when the result is not significant. Experimentation skills are in short supply and easy to check in an interview.
4. A small LLM or agent feature with evaluation. In 2026 the bar is not "I called an API." It is "I built a retrieval pipeline, wrote twenty test cases, measured where it fails and fixed two of them." Show the evaluation table, not just the demo.
What gets you filtered out
- Five half-finished notebooks instead of one finished project.
- No stated question — just a dataset and a wall of charts.
- Accuracy reported on the training set, or no baseline to compare against.
- Code that only runs on your laptop. Include a requirements file and a one-command way to reproduce the headline result.
- A README that describes the tools instead of the outcome.
How to present it
One paragraph summary at the top, one chart that carries the result, a short "what I would do next" section that shows judgment. Add a two-minute Loom walkthrough if you can; it doubles as interview practice. In the interview itself, be ready to defend one decision you made and one you would change.
Where to get projects like these
The graded assignments in the free academy are designed to become portfolio pieces — each one ends with a deliverable, not a quiz. Interns in the Mentor-Led Internship work on real client problems with a mentor reviewing every week, which is why their portfolios pass the ninety-second test. Either way: pick one project, finish it properly, and write about what you found.