About StaffML
9,438 physics-grounded ML systems questions across 87 topics and 4 deployment tracks, at 6 difficulty levels from recall to system design. Backed by the Machine Learning Systems textbook. Free, open source, and runs entirely in your browser.
Read the paper
StaffML is described in a research paper on corpus design and competency-backed question authoring for ML systems.
vpreview-dev · May 5, 2026 · hash 04ee8a2
▸Cite this release
@misc{staffml2026,
title = {StaffML: A Physics-Grounded Interview Question Bank for Machine Learning Systems Engineers},
author = {Janapa Reddi, Vijay and contributors},
year = {2026},
version = {vpreview-dev},
note = {Release hash: 04ee8a2322b7f531},
url = {https://staffml.mlsysbook.ai}
}release_hash: 04ee8a2322b7f531fa528b7b323f3e7a4c3ac17e55f0643a44fadb00fc2183fb
“Every semester, students come to my office hours with the same question: how do I prepare for ML interviews? Not the modeling side — the infrastructure. The compute, the memory, the hardware, the deployment. These interviews expect you to reason about all of it, often with numbers, on the spot. And most people aren't ready.
That's why I built StaffML. I wanted to give students and engineers a way to find out what they really know, and what they still need to learn. The questions come straight from the Machine Learning Systems textbook, but a textbook teaches you concepts one at a time, and quizzes test whether you remember them. Interviews challenge you to connect concepts across the entire system stack. StaffML does just that.
It's free because interview prep is just another form of education, and education works best when it's free and open access, and the world needs more AI engineers.”
What makes StaffML different
Every question traces back to a specific chapter of the Machine Learning Systems textbook. You are learning the 87 concepts a curriculum designer chose, not whatever the internet happened to cough up.
When a question asks about memory bandwidth, the numbers come from actual H100, A100, and Jetson datasheets. The math works on real silicon, not on round numbers that make the answer tidy.
Questions ask you to estimate, diagnose, compare tradeoffs, and architect — the same skills tested in Staff+ interview loops. Nothing asks you to recite a definition.
Type your calculation, then compare it against the model answer. The app tells you whether you are in the right ballpark or off by a factor of ten.
Every question passed a second-pass math check by a separate model. The initial verification pass flagged an 8.3% error rate across the corpus. All flagged errors were corrected.
AI is not magic — it is infrastructure, and infrastructure has laws.
StaffML is part of the Machine Learning Systems curriculum at Harvard University. Every topic links back to its source chapter.
Try a Question
Who is this for?
Start with L1–L2 recall questions to build your foundation.
Jump to L4–L6+ questions. Try the Mock Interview.
Do the Daily Challenge — 3 questions, 5 minutes, same for everyone.
Browse the question bank to see what the field looks like.
How Questions Are Organized
Every question is tagged by difficulty (6 levels from recall to system design), competency zone (what kind of thinking it tests), and deployment track (Cloud, Edge, Mobile, or TinyML).
Difficulty Levels
“What does GPU HBM stand for?”
“Why is memory bandwidth often more important than FLOPS for inference?”
“Calculate the minimum batch size to saturate an H100's memory bandwidth.”
“Your serving latency spiked 3x after switching FP16→FP8. Why?”
“Design a serving stack for Llama-70B at 10K QPS on a $50K/month budget.”
“Design fault-tolerant training for a 1T param model across 3 data centers.”
Deployment Tracks
How Questions Are Built
StaffML questions are constructed using LLM-assisted generation with structured prompts grounded in the Machine Learning Systems textbook and the MLSysIM physics engine. Every hardware specification traces back to a centralized constants table maintained alongside the textbook.
Every question undergoes independent math verification by a separate model that rechecks all arithmetic and hardware specs. The initial verification pass found an 8.3% error rate across the corpus. All identified errors were corrected.
The full methodology — backward design from textbook chapters, four-axis taxonomy, LLM-assisted generation pipeline, independent math verification, and the ikigai-inspired competency zone framework — is described in our paper.
Found an error? We take correctness seriously. If you spot a wrong number, a broken calculation, or a misleading scenario, open an issue on GitHub. Community verification is how we keep improving.
Open Source
The entire question corpus, taxonomy, and web application are open source. Contributions, feedback, and corrections are welcome.
View on GitHubvpreview-dev · built 2026-05-05