Skip to content
mentorship

Free prep resources

Choose hours, not just weeks.

A plan is executable only when it states the weekly workload, recurring practice lanes, outputs, and conditions for extending or stopping. The schedules below assume eight hours per week and show how to scale.

2-week planFast consolidation · 16 standard hours4-week planFocused rebuild · 32 standard hours8-week planBreadth plus depth · 64 standard hours

Start from evidence

Use the readiness and workload check. Two weeks is for consolidation, four for bounded repair, and eight for transitions or several critical gaps. If your runway is shorter than the evidence suggests, triage to actual rounds or consider moving the loop.

Weekly workload contract

CapacityNew attemptsLong roundsRetriesStoriesRule
5h · Minimum1 short1 alternating long round21 rehearsalReduce new topics first; keep retries and simulation.
8h · Standard2 short1 long round31–2 rehearsalsThe weekly schedules below assume this workload.
12h · Intensive3 short2 long rounds42 rehearsalsAdd depth and feedback, not indiscriminate topic coverage.

A “short attempt” includes the timed answer, rubric diagnosis, targeted reading, and a rewritten outline, usually 45–60 minutes. A “long round” is a 35–45 minute ML implementation, design, research, experiment, or project session plus review.

Standard eight-hour weekly template

Mon · 60m

New short attempt

Use one high-priority question. Score before reading.

Tue · 45m

Due retry

Change the follow-up; repair one observable weak dimension.

Wed · 90m

Full long round

Rotate ML implementation, system design, research, experiment, or project deep-dive by role.

Thu · 60m

New short attempt

Interleave a different category from Monday.

Fri · 45m

Due retry

Test whether earlier repair survived spacing.

Sat · 90m

Story + pushback

One 90-second opening, three follow-ups, and one recording review.

Sun · 60m

Mixed session

Random prompt or transfer drill with no category cue.

Sun · 30m

Review and flex

Update due dates, drop optional topics, and protect spillover capacity.

Role overlays

Apply the allocation before choosing questions. The same four-week calendar should not look identical for AS, MLE, and RE candidates.

Applied Scientist

Balance scientific depth with scoping, product judgment, and evidence that you can ship.

  • 25% ML breadth & research depth
  • 30% ML design & experimentation
  • 20% Project deep-dive
  • 15% Behavioral & leadership
  • 10% Coding / implementation

Do not prep like a pure researcher: shipping evidence, product metrics, and engineering judgment often decide the loop.

Five things AS interviews test · Evaluate an LLM application · Most ambitious project

Machine Learning Engineer

Prioritize coding and dependable systems while retaining enough ML depth to make sound modeling decisions.

  • 30% ML implementation / coding
  • 25% ML system design
  • 20% Production & infrastructure
  • 15% ML breadth
  • 10% Behavioral & project evidence

Do not substitute ML reading for coding repetitions. Many MLE loops reject strong modelers on software execution.

Debug a training loop · Design ML monitoring · Design a feature store

Research Engineer

Combine strong implementation and systems reasoning with enough research depth to critique and operationalize ideas.

  • 30% ML implementation / coding
  • 25% Training & inference systems
  • 20% Research depth
  • 15% ML breadth
  • 10% Project & behavioral evidence

Expect title variation. Confirm whether the loop behaves like research, ML systems, or general software engineering.

Implement attention · Train a 100B model · Design an ablation study

Target-level evidence

L4 / mid-level

Bar: Execute a bounded problem correctly and explain the technical choices.

  • Correct implementation
  • Sound fundamentals
  • Clear debugging procedure
  • Awareness of basic trade-offs

Common failure: Trying to manufacture strategy instead of demonstrating reliable execution.

L5 / senior

Bar: Own an ambiguous project area and make autonomous, defensible decisions.

  • A clear personal decision thread
  • End-to-end shipping evidence
  • Trade-offs and failed attempts
  • Online or operational outcomes

Common failure: Describing team activity without identifying the decisions you personally owned.

L6 / staff

Bar: Choose problems and strategy, influence multiple teams, and remain technically credible.

  • Cross-team adoption or influence
  • A wrong strategic bet and recovery
  • A project you killed or redirected
  • Hands-on depth beneath the strategy

Common failure: Speaking only at the strategic level without concrete technical decisions and failure evidence.

Fast consolidation · 16 standard hours

2-week plan

Best for: A candidate whose important rounds already produce workable timed answers.

This is not a foundation-building plan. If three critical areas are at 0–2 evidence, use four or eight weeks when possible.

Week 1 · Diagnose and repair

Prove the baseline, choose three gaps, and begin delayed retrieval.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 2 × 45 min: timed short-form baselines from different critical rounds, including rubric review
  • 1 × 90 min: one full-length ML implementation, design, research, or project deep-dive baseline
  • 2 × 45 min: delayed retries of the weakest dimensions, at least 48 hours later
  • 1 × 90 min: draft six story outlines; rehearse the two most level-relevant stories
  • 1 × 60 min: repair only mechanisms or evidence exposed by attempts
  • 1 × 30 min: update the retry queue and schedule Week 2
  • 1 × 30 min: flex, spillover, or an extra due retry

Required output: Three bounded gaps, six story outlines, one long-round baseline, and every Weak result scheduled.

Exit gate: Each critical round has a recent baseline. If any round is still unattempted, do not call the plan “on track.”

Week 2 · Simulate, repair, taper

Integrate the loop early enough to repair evidence, not feelings.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 1 × 210 min: full role-specific simulation, including short breaks; complete it early
  • 2 × 45 min: repair the two highest-impact simulation failures
  • 2 × 30 min: final delayed retries; no answer graduates after one good repetition
  • 1 × 45 min: rehearse three stories in 90–120 second form with skeptical follow-ups
  • 1 × 30 min: concise outlines, formulas, and known failure points only
  • 1 × 45 min: logistics, spillover, and recovery; no new topic sprint

Required output: One scored simulation, two repaired failures, and a final list of known (not hidden) risks.

Exit gate: Proceed confidently only if every critical round has a workable recent attempt and the simulation exposes no unbounded failure.

Focused rebuild · 32 standard hours

4-week plan

Best for: A candidate with a foundation but inconsistent retrieval or a small number of important gaps.

Do not spend four weeks reading. Every week contains new attempts, delayed retries, one long round, and behavioral/project evidence.

Week 1 · Baseline the actual loop

Replace confidence estimates with timed evidence.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 2 short-form baselines from distinct rounds
  • 1 full-length baseline from the highest-weight long round
  • 3 delayed retries or follow-up drills
  • 2 story outlines plus one spoken rehearsal
  • Targeted concept repair, weekly review, and flex

Required output: A ranked gap list, evidence for every expected round, and a role-specific allocation.

Exit gate: No “unknown” critical round. Untested is a result and remains the highest risk.

Week 2 · Build depth without dropping other lanes

Make the primary domain follow-up ready while continuing ML implementation, design, and story repetitions.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 2 domain-specific short attempts with one unfamiliar follow-up each
  • 1 long round selected by role: ML implementation, system design, research, or experimentation
  • 3 retries from Week 1 before adding optional topics
  • 1 project or behavioral rehearsal with ownership partition
  • One mechanism-level repair block, weekly review, and flex

Required output: One deep answer per primary technical domain and a shrinking (not merely changing) retry queue.

Exit gate: A strong headline answer that fails the first “why?” follow-up is not ready.

Week 3 · Interleave systems, decisions, and evidence

Remove predictable cues and pressure-test senior judgment.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 1 short breadth attempt selected at random
  • 1 product/experiment or transfer drill in an unfamiliar domain
  • 1 full ML implementation, system design, research, or project deep-dive round
  • 3 delayed retries, including at least one previously improved answer
  • 2 behavioral stories under skeptical pushback
  • Weekly review and flex

Required output: Mixed-round performance, four rehearsed stories, and evidence that prior repairs survived spacing.

Exit gate: At least 70% of critical rubric dimensions are solid across a mixed session; no single round is consistently Weak.

Week 4 · Simulate and consolidate

Expose integrated failure early, repair narrowly, then protect execution.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 1 full role-specific simulation in the first half of the week
  • 2 bounded repairs from simulation evidence
  • 2 final delayed retries
  • 2 concise story rehearsals plus role narrative
  • Final-week checklist, logistics, taper, and recovery

Required output: A scored loop, an explicit risk decision, and no surprise unattempted round.

Exit gate: If a critical round remains Weak after repair, extend or move the loop when possible; do not hide it with more reading.

Breadth plus depth · 64 standard hours

8-week plan

Best for: Role transitions, broad loops, several critical gaps, or candidates rebuilding retrieval from passive knowledge.

The focus changes by week, but the lanes do not. Every week maintains retrieval, a long round, project/behavioral evidence, and review.

Week 1 · Diagnose and bound

Baseline actual rounds; choose three priorities; start six stories.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 2 short baselines
  • 1 long-round baseline
  • 2 first retries
  • 2 story outlines and one rehearsal
  • Role/level calibration, review, and flex

Required output: Round map, three priorities, six story candidates.

Exit gate: Every critical round is either attempted or explicitly scheduled for Week 2.

Week 2 · Repair foundations

Fix mechanisms that block higher-level reasoning while maintaining delivery practice.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 2 fundamentals or math attempts
  • 1 role-specific long round
  • 3 Week 1 retries
  • 1 project/behavioral rehearsal
  • Targeted concept repair, review, and flex

Required output: Mechanism-level answers for the highest-impact prerequisites.

Exit gate: You can explain why the method works and one condition where it fails.

Week 3 · Primary domain depth

Connect mechanisms, metrics, failure modes, and operational decisions.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 2 primary-domain attempts
  • 1 long ML implementation, design, or research round
  • 3 spaced retries
  • 1 story rehearsal with technical follow-ups
  • Reference repair, review, and flex

Required output: Two follow-up-ready domain answers and one defensible project decision.

Exit gate: One unfamiliar follow-up does not collapse the answer into memorized phrases.

Week 4 · Design and experimentation

Build end-to-end judgment from objective through launch and monitoring.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 1 ML design attempt
  • 1 experimentation or transfer drill
  • 1 role-specific long round
  • 3 retries
  • 1 behavioral rehearsal
  • Review and flex

Required output: A coherent architecture and an explicit ship/iterate/stop rule.

Exit gate: The answer links offline evidence, online outcomes, guardrails, cost, and rollback.

Week 5 · Production and implementation

Turn models into reliable systems while preserving all other lanes.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 1 debugging or production attempt
  • 1 short breadth attempt
  • 1 full ML implementation or infrastructure round
  • 3 retries
  • 1 project story about failure/recovery
  • Review and flex

Required output: A systematic debugging procedure and a tested implementation under time.

Exit gate: You can identify the bottleneck or next diagnostic before proposing random fixes.

Week 6 · Project and leadership depth

Pressure-test level, ownership, influence, wrong bets, and killed work.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 1 short technical maintenance attempt
  • 1 long project deep-dive
  • 3 retries
  • 3 behavioral stories in 90–120 seconds
  • One instructed pushback drill
  • Review and flex

Required output: Six complete stories and one 20-minute project deep-dive.

Exit gate: Ownership is bounded, target-level scope is visible, and skeptical follow-ups reveal more (not less) specificity.

Week 7 · Mixed simulation and repair

Remove domain blocks and test the integrated loop.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 1 full role-specific simulation
  • 2 high-impact repairs
  • 2 delayed retries
  • 1 recorded story review
  • One short random transfer drill
  • Risk review and flex

Required output: A scored simulation and no more than three bounded repair tasks.

Exit gate: If more than one critical round remains Weak, extend the plan when possible.

Week 8 · Consolidate and taper

Finish evidence, avoid new-topic overload, and protect interview execution.

Standard workload: about 8 hours, including review and flex.

Standard 8-hour blocks

  • 2 final due retries
  • 1 short mixed confidence session
  • 2 concise story rehearsals
  • Role narrative and interviewer questions
  • Final-week checklist, logistics, rest, and flex

Required output: Known risks, stable retrieval, tested logistics, and preserved energy.

Exit gate: No critical round is unattempted; every graduated gap has two spaced successes.

Readiness and extension gates

Graduate one gap

  • Two successful closed-book attempts separated in time
  • No critical rubric dimension still marked weak
  • One unfamiliar follow-up handled without notes
  • The answer survives a later mixed session

Call the loop ready

  • Every expected round has a recent full attempt
  • No critical round remains consistently Weak
  • One role-specific simulation completed under realistic timing
  • Six stories are concise, specific, and defensible

Extend or move when possible

  • A critical round remains unattempted one week before the loop
  • More than one critical round remains Weak after simulation repair
  • ML implementation or design cannot finish a correct baseline in the round time
  • Stories lose ownership or specificity under follow-up

If you fall behind

Remove optional new topics first. Then reduce the number of new attempts. Preserve high-priority retries, one realistic simulation, behavioral/project evidence, and recovery. A smaller set of durable answers is better than shallow completion.