Big Tech platforms Applied Scientist (AI/ML) Interview Loop
Applied Scientist is a title Amazon and Microsoft use. Meta and Google call the closest role Research Scientist and run a different loop. So this page covers the Amazon and Microsoft applied scientist interview. At Amazon, there are one or two phone screens with a senior scientist. Then come four to six hour-long rounds: science breadth, science depth, live coding, science application or system design, and the Bar Raiser. At Microsoft, a 30-minute screen mixes ML coding, a project deep dive and fundamentals. Then there are three to five rounds, built around a hiring manager and two design conversations.
Companies behind this pattern
Practice by round
A medium coding problem in Python, plus breadth questions. Expect bias and variance, loss functions, optimization, and how you'd fix a model that overfits a small, noisy dataset. Many teams also ask SQL on their own data, because in this role you pull your own features.
Live coding at loop level, without an IDE and sometimes with an ML angle. There's also the SQL and Python data handling that sits under any science application.
| Question | Difficulty |
|---|---|
| Meta/Facebook Accounts | Medium |
| Array Into Bins | Medium |
| Randomized Distribution Generator | Medium |
| Bernoulli Matrix Column Normalization | Medium |
| Linear Fit Algorithm | Medium |
| Best Selling Item | Hard |
| Popularity Percentage | Hard |
| Matrix Multiplication | Hard |
| Bipartite Graph Detection | Hard |
| Nested JSON Flattening | Hard |
| Longest Path in DAG | Hard |
| Maximum Path-Sum in Tree | Hard |
| Binary Tree Traversal | Hard |
| Graph Topological Sort | Hard |
| Largest Perimeter | Hard |
Science application and system design. You take a business problem, like fraud detection, a recommender, feed ranking or a customer-service quality metric, and turn it into a system. That means the data, the model, the evaluation and the serving. In the real loop, the depth round takes a past project apart, one decision at a time. The rating-prediction and clustering projects here are practice for that. The most common rejection is a candidate who can't put a number on the impact of their own work.
Science breadth: how boosting works, random forests, regularization, and few labeled instances. LLM teams add Transformer attention complexity and the KV cache. On the statistics side, there are sampling distributions and A/B testing.
| Question | Difficulty |
|---|---|
| Mean and Median | Easy |
| Calculating F1 Score | Easy |
| Broken Packages | Medium |
| A/B Testing | Medium |
| Multi-dimensional Data | Medium |
| Statistical Hypothesis | Medium |
| Skewed Data | Medium |
| Difference-in-Differences for a Country Launch | Medium |
| Propensity Matching for an Opt-In Feature | Medium |
| Sampling Distribution and Type 1 Errors | Hard |
| Mean, Median and Mode | Hard |
| An Instrument From a Rollout Glitch | Hard |