Skip to content
mentorship

Free prep resources

Practice the work sample, not just the answer.

Current senior ML and Research Engineer loops increasingly use codebases, traces, notebooks, presentations, and open-ended investigations. These labs turn those formats into bounded, executable practice without publishing leaked questions.

Use only confirmed formats

Do not complete every lab. Ask the recruiter for the assessment type, duration, editor, AI policy, expected artifact, and review format. Then choose the closest exercise and keep the rest optional.

AI-native engineering work sample

Agentic ML codebase lab

60 minutes Extend an unfamiliar ML evaluation package with an AI coding agent while preserving correctness, scope, and ownership.

LLM training debugging work sample

Broken frontier LLM training lab

75 minutes Diagnose interacting data, masking, loss, accumulation, scheduler, and evaluation defects in a tiny causal LM.

Research work sample

Black-box model behavior lab

90 minutes Turn model observations into falsifiable hypotheses, discriminating probes, boundary conditions, and a short research readout.

ML primitives in code

Frontier ML implementation set

Five sessions, 35 to 50 minutes each Implement a decoder block, incremental KV cache, beam search, LoRA adapter, and reverse-mode autodiff against executable contracts.

Inference systems work sample

LLM inference scheduler lab

75 minutes plus 30-minute design follow-up Implement KV admission, chunked prefill, continuous batching, tenant fairness, and overload behavior for a model server.

Performance engineering work sample

Accelerator performance lab

Two hours for a timed attempt, then open-ended Use a trace, bottleneck model, and correctness-preserving experiments to optimize Anthropic’s released simulated-accelerator challenge.

Math and statistics screen

Timed ML math oral

Eight separate drills, 8 to 15 minutes each Derive, sanity-check, and interpret eight ML results under oral follow-up rather than silently reproducing memorized algebra.

Post-training research work sample

Post-training environment and grader lab

60 minutes plus 20-minute research discussion Design an RL environment and repair a grader that rewards successful but unsafe tool use.

Rules shared by every lab

  1. Start from the contract. State inputs, outputs, invariants, timebox, and forbidden shortcuts before touching code.
  2. Preserve evidence. Keep failing tests, traces, probe results, and rejected hypotheses. A clean final state hides how you reasoned.
  3. Use AI only when allowed. Some labs intentionally support an agent. Others model closed-book or no-AI rounds.
  4. Review before accepting. Generated code, grader outputs, and model explanations are untrusted until you verify them.
  5. Finish with a decision. State what works, what remains uncertain, and the next highest-information action.

Not a question dump

The labs are original exercises derived from public format evidence and current role demands. They do not reproduce confidential prompts, and they should not be presented as exact company questions.

Check current process evidence →