Capstone Project & Interview Prep
Turn what you have learned into a portfolio project that proves it, and prepare for each stage of an ML engineering interview.
What employers look for in a junior ML engineer
Hiring managers for junior ML roles rarely expect deep research expertise. They look for evidence of four things. Solid fundamentals: you can explain bias and variance, leakage, regularisation, evaluation metrics and how common models work, without hand-waving. Practical skills: Python, pandas, SQL, scikit-learn and at least one deep-learning framework, plus git and the ability to write clean, tested code. Engineering judgement: honest validation, sensible baselines, knowing when a simple model is enough, thinking about deployment, monitoring and cost. And communication: you can explain a result to a non-technical person and push back politely when something is wrong.
The best way to show all four at once is a portfolio project that looks like real work: a messy dataset, a clear problem, honest evaluation, a deployed or at least packaged model, and a write-up that leads with the outcome. One excellent project is worth more than ten tutorial notebooks.
The capstone project
The capstone repeats the full workflow from the game on a problem you choose. Pick a problem with a real decision behind it ("which customers should get a retention call?", not just "predict column X"), a public dataset with some mess in it, and a scope you can finish in four to six weeks.
| Week | Deliverable |
|---|---|
| 1 | Problem statement: target, prediction time, action, success metric, constraints. Data audit: issues, leakage risks, privacy notes. |
| 2 | Data pipeline in code (SQL and/or pandas, tested). Time-aware train/validation/test split. Baselines. |
| 3 | Model bake-off with cross-validation, tuning, error analysis. Threshold chosen from costs. |
| 4 | Responsible AI: per-group metrics, explanations, a model card. |
| 5 | Packaging: saved pipeline, a predict() function or small API, input validation, a simple monitoring check. |
| 6 | Write-up: README that leads with the result, a short demo (screenshots, a notebook or a small web app), and what you would do next. |
Put it in a public GitHub repository with a clear structure (data/, src/, notebooks/, tests/, README.md, requirements.txt), commit as you go, and make sure someone else can reproduce your main result by following the README.
Try it: plan your capstone
Choose a domain and the planner suggests datasets, a decision to support, likely pitfalls and a week-by-week checklist. Tick items off as you go; your progress is saved in this browser.
Interactive: this part runs in your browser.
How ML interviews work
Most ML engineering interview loops have four to five stages. A recruiter screen checks motivation, experience and logistics. A technical screen asks rapid ML fundamentals ("explain overfitting", "precision vs recall", "how does a random forest reduce variance?"). A coding round tests Python, data manipulation (pandas or SQL) and sometimes implementing a metric or a simple algorithm from scratch. An ML system design round asks you to design an end-to-end system. A behavioural round explores how you work with others. A take-home project sometimes replaces the coding round.
For ML system design, use a consistent structure and say it out loud: (1) clarify the goal and the decision, and agree on a business metric; (2) frame it as an ML problem (target, prediction time, labels and how they arrive); (3) data and features, with leakage and privacy risks; (4) a baseline, then the model, and why; (5) offline evaluation (split, metric, threshold); (6) serving (batch or real-time, latency, scale); (7) monitoring, feedback loops and retraining; (8) risks: fairness, safety, failure modes. Interviewers care more about structured trade-offs than about naming the fanciest model.
For behavioural questions, answer with STAR: the Situation, your Task, the Actions you personally took, and the Result (with a number if possible) plus what you learned. Prepare five stories: a technical challenge, a disagreement, a mistake you fixed, working under a deadline, and explaining something complex to a non-expert. The game's milestones are good material, and so is your capstone.
Try it: interview flashcards
Thirty common ML interview questions with model answers. Say your answer out loud first, then flip the card and rate yourself. Cards you find hard come back more often.
Interactive: this part runs in your browser.
Code example: a classic coding-round question
A worked example: read it step by step (every part is explained), then run it and experiment.
Interactive: this part runs in your browser.
Advanced: your first 90 days and beyond
Once hired, the habits from the game matter most. Ask about the decision before the model. Write things down: problem statements, design docs, experiment logs, postmortems. Ship small: a baseline in production teaches more than a perfect model on a laptop. Make your work reproducible and reviewed. Learn the domain from the people who do the work. And keep a "brag document" of what you shipped and its impact; it will make performance reviews and your next interview much easier.
To grow towards mid-level, deepen one area (for example MLOps, NLP and LLM systems, recommender systems, or forecasting) while staying broad, read your team's production code, volunteer for on-call and incident reviews, and mentor someone newer than you. For research-oriented paths, add reading and reproducing papers, rigorous experimental design and writing up results clearly.
Quiz
1. In STAR, what does the R stand for, and what should it include?
2. In a system-design interview, what should you do first?
3. Which portfolio is strongest for a junior ML role?
4. Your capstone README should start withโฆ
๐ Further reading
Trusted, mostly free sources: official documentation, well-known textbooks and courses. When you want to confirm or go deeper on anything here, start with these.
- Google: Rules of Machine Learning โ Best practices for ML engineering
- Fairness and Machine Learning (free book) โ Barocas, Hardt, Narayanan