Amazon Applied Scientist (AI/ML) Interview Loop
Amazon's applied scientist loop starts with a recruiter call, then one or two technical screens with a senior scientist. The screen mixes questions about your own projects, ML fundamentals and one coding problem. The onsite is four to six interviews. Most loops include a short research talk, an ML breadth round and a deep round on your own work. Then an ML design problem, coding and a Bar Raiser. Every interviewer also asks a few Leadership Principles questions.
What decides it
Depth on your own work and judgment on open problems. Coding needs to be solid but it is rarely the reason for a no. People fail by naming methods without saying when they would use them, and by telling Leadership Principles stories without a clear result.
Practice by round
Quick, clean SQL and Python. Amazon rarely asks SQL in this loop. The coding screen still expects you to handle data without effort.
Coding at onsite level. Expect one problem on trees, graphs or dynamic programming, written in real syntax. Some teams add an ML routine coded from scratch, like computing AUC or sampling from a distribution.
| Question | Difficulty |
|---|---|
| Monthly Sales Rolling Average | Medium |
| Finding User Purchases | Medium |
| Player with Longest Streak | Hard |
| Popular Posts | Hard |
| Cookbook Recipes | Hard |
| First Day Retention Rate | Hard |
| Sales Growth per Territory | Hard |
| Product Market Share | Hard |
| Customer Tracking | Hard |
| Matrix Multiplication | Hard |
| Nested JSON Flattening | Hard |
| Binary Tree Traversal | Hard |
| Graph Topological Sort | Hard |
| Multi-Status Order Tracking by Week | Hard |
| Daily Revenue | Hard |
The applied science problems: predicting which deals will fail, estimating grocery delivery arrival times, designing a click-through estimation system, handling severe imbalance in fraud flags, and choosing a model under constraints. Each is an open problem. The interviewer wants to see you frame it, pick a baseline and say how you would know it works. There are also two data projects, seller return rates and deal success rate, to work end to end.
| Case study | Difficulty |
|---|---|
| Predicting Which Deals Will Fail | Medium |
| Severe Imbalance in Fraud Flags | Medium |
| Choosing a Model Under Constraints | Medium |
| Grocery Delivery Arrival-Time Prediction | Hard |
| Designing a Click-Through Estimation System | Hard |
| Data project | Difficulty |
|---|---|
| Seller Return Rates | Medium |
| Deal Success Rate | Medium |
Statistics and probability with follow-ups: bias and variance, reading a p-value, which carrier delivered a late parcel, the first one-star review, and maximum likelihood on coin flips. Teams working on language models also ask about attention and fine-tuning, so have an answer ready even though it is not practiced here.
| Question | Difficulty |
|---|---|
| Random Bucketing | Easy |
| Interpret P-value | Easy |
| Assessing Multicollinearity | Easy |
| Bias-Variance Tradeoff | Medium |
| Linear Regression & T-test | Medium |
| Broken Packages | Medium |
| Multi-dimensional Data | Medium |
| Statistical Hypothesis | Medium |
| Which Carrier Delivered Late | Medium |
| First One-Star Review | Medium |
| Maximum Likelihood of Heads | Hard |
| An Instrument From a Rollout Glitch | Hard |