Before choosing a path
Company titles are inconsistent. Ask for round formats, implementation style, project expectations, and the day-to-day research/modeling/engineering/product mix. If the recruiter cannot provide detail, use the typical rounds in the readiness check and mark uncertainty as risk.
Ordered core path
Applied Scientist
Do not prep like a pure researcher. Preserve science, but prove scoping, experimentation, shipping, and influence.01 · Calibrate role and level
Read the role taxonomy and L5/L6 calibration; choose the target rubric.
02 · Scope before solving
Practice objective, user, data, constraints, and smallest discriminating experiment.
03 · Prove ML breadth
Explain mechanisms and alternatives, not recipes.
04 · Design end to end
Connect data, baseline, model, eval, serving, monitoring, and iteration.
05 · Make a product decision
Choose metrics, experiment design, validity checks, and a ship rule.
06 · Debug evidence disagreement
Separate objective mismatch, leakage, serving skew, and feedback effects.
07 · Defend project ownership
Prepare a short opening and 20-minute technical deep-dive.
08 · Show judgment under pressure
Balance legitimate delivery pressure with measurable user risk.
09 · Simulate the loop
Run the AS packet and reduce results to three bounded repairs.
Ordered core path
Machine Learning Engineer
Do not substitute ML reading for software execution. Code, debug, and design reliable systems repeatedly.01 · Confirm the loop
Ask whether implementation means general algorithms, ML primitives, debugging, or all three. Use other resources for general algorithms.
02 · Establish an ML implementation baseline
Build mergeable model-evaluation logic with tests, edge cases, and explicit metric semantics.
03 · Add bounded-memory implementation
Reason about vectorization, memory, and exact-versus-ANN trade-offs.
04 · Debug systematically
State the diagnostic order before editing code.
05 · Design reliable serving
Cover data, model, outcome, alerts, ownership, and rollback.
06 · Make cost explicit
Quantify workload and quality/cost/latency frontiers.
07 · Debug offline and online disagreement
Separate experiment integrity, serving skew, proxy mismatch, and feedback effects.
08 · Prepare production failure and recovery stories
Use failure, incident recovery, trade-off, and cross-team stories.
09 · Simulate the loop
Run ML implementation, design, production, breadth, and project rounds.
Ordered core path
Research Engineer
Balance implementation and systems depth with scientific skepticism. Confirm whether the title is research-heavy or software-heavy.01 · Confirm role mix
Map research, modeling, engineering, and product expectations.
02 · Implement a core primitive
Build the mechanism correctly and explain dimensions and complexity.
03 · Add systems implementation
Handle memory and performance without hiding behind a library.
04 · Model the workload
Reason about memory, parallelism, communication, failures, and checkpoints.
05 · Reason about inference
Use measurement and cost models rather than a list of optimizations.
06 · Design discriminating evidence
Control compute, tuning, variance, and alternative explanations.
07 · Critique research fairly
Prioritize the threat that most changes the central claim.
08 · Prepare translation stories
Show failed experiments, implementation decisions, and research-to-production impact.
09 · Simulate the loop
Run ML implementation, systems, research depth, breadth, and project rounds.
Frontier format overlays
Do not add these because the company is fashionable. Add one only when recruiting confirms the format.
Agentic codebase
Use when: The recruiter names an existing repository or authorized coding agent.
Replace one implementation repetition with the agentic evaluation-service lab. Practice mapping, bounded delegation, diff review, and proof.
Open practice →Technical presentation
Use when: The loop includes a project presentation, job talk, or artifact defense.
Prepare 30 minutes around three decisions and reserve 15 minutes for interrupted questions.
Open practice →Research work sample
Use when: The assessment uses a notebook, API, paper, model, or open-ended investigation.
Run the black-box lab and produce a five-minute claim, evidence, uncertainty, and next experiment readout.
Open practice →Values and mission
Use when: The loop includes a dedicated values, safety, mission, or senior-leadership discussion.
Prepare four real tensions and practice independent judgment under principled pushback.
Open practice →Performance trace
Use when: The role centers on kernels, accelerators, training throughput, inference, or systems performance.
Use the released accelerator challenge or inference scheduler and keep an experiment ledger.
Open practice →Domain supplements
A domain supplement changes examples and technical depth. It does not replace ML implementation, project evidence, behavioral judgment, or the actual round formats.
LLM and agents
- LLM evals and failure taxonomy
- RAG versus fine-tuning decisions
- Inference workload and cost
- Agent reliability and tool-use evaluation
Evaluate an LLM application · Inference service · Evaluate an agent
Post-training and environments
- Preference data and reward models
- RL environments and evidence-bearing graders
- Synthetic data and held-out behavioral families
- Reward hacking and online product validation
Alignment and model behavior
- Scalable oversight and control
- Mechanistic interpretability
- Monitorability and model behavior
- Adaptive red-team and LLM security
Scalable oversight · Mechanistic interpretability · Red-team design
Multimodal and robotics
- Cross-modal alignment and fusion
- Image, audio, and video token budgets
- Grounding and modality-conflict evaluation
- Observation, action, safety, and sim-to-real
Recommendations and search
- Candidate generation and ranking
- Offline ranking metrics and position bias
- Exploration and feedback loops
- Cold start and ecosystem health
ML platform and infrastructure
- Training and serving workload models
- Distributed communication and memory
- Feature/data reliability
- Monitoring, fallback, and cost ownership
General product ML
- Problem framing and baseline
- Experiment and metric hierarchy
- Delayed labels and feedback
- Cross-functional decision evidence
Choose product metrics · Design an A/B test · Fraud detection
Research and modeling
- Hypothesis and discriminating evidence
- Matched baselines and compute
- Ablations and alternative explanations
- Derivations and boundary conditions
Design an ablation · Critique a paper · Derive logistic regression
Specialist-domain boundary
The core library is deepest in LLMs, recommendations, general ML, training, and systems. For specialized CV, speech, or RL loops, use the concept collections below for foundations, but obtain a domain-specific paper/project mock from someone current in that field.
What the core library does not yet cover deeply
This site does not attempt to replace general algorithms or SQL interview preparation. The implementation questions stay here only when the task is fundamentally about ML: attention, training-loop diagnosis, memory-bounded retrieval, or mergeable evaluation metrics. Use external, current material for general data structures, framework internals, streaming-data infrastructure, privacy/compliance implementation, advanced causal inference, or a specialist research frontier.
How to use the sequence
Do not merely read from 01 to 09. Attempt each linked question first, diagnose with its round-specific rubric, schedule retries, and move forward while due items continue in parallel. Run the simulation only after every expected round has a baseline.