Skip to content
mentorship

Free prep resources

Practice the round you will actually face.

An eight-minute verbal answer, a forty-minute design, and an ML implementation task are different performances. Question pages now choose a timer and rubric based on the round rather than applying one generic scorecard.

Use Practice Mode on any question page. Scratchpad text stays in memory. If enabled by you, the local queue stores only question ID, round mode, score, weak rubric dimensions, attempt count, and due date.

Round modes

ML breadth

8 min

Explain the mechanism, why it works, when it fails, and one alternative.

  • Mechanism
  • Assumptions
  • Trade-offs
  • Follow-up depth
  • Communication

ML implementation

40 min

Clarify the contract, implement a correct baseline, test edge cases, then optimize.

  • Contract & examples
  • Correctness
  • Complexity
  • Testing & debugging
  • Communication

ML system design

40 min

Scope first; then cover data, baseline, model, evaluation, serving, monitoring, and iteration.

  • Problem framing
  • End-to-end architecture
  • Evaluation & decision
  • Operations & failure
  • Communication

Project deep-dive

20 min

Give a two-minute overview, then pressure-test decisions, failures, evidence, and ownership.

  • Scope & stakes
  • Personal ownership
  • Decision quality
  • Failure & recovery
  • Impact & reflection

Behavioral / leadership

3 min

Answer the exact question in 90–120 seconds, then reserve detail for follow-ups.

  • Question fit
  • Ownership
  • Stakes & conflict
  • Evidence & result
  • Reflection

Product & experimentation

20 min

Define the decision first, then design the measurement and enumerate validity threats.

  • Problem framing
  • Metrics
  • Experiment design
  • Decision rule
  • Communication

Research depth

25 min

State the claim, test its evidence, design discriminating experiments, and identify alternatives.

  • Claim & hypothesis
  • Evidence quality
  • Ablations
  • Limitations & transfer
  • Communication

Systems / infrastructure

35 min

Quantify the workload, identify the bottleneck, choose an architecture, and design failure recovery.

  • Workload model
  • Bottleneck reasoning
  • Architecture & trade-offs
  • Reliability
  • Communication

Math / derivation

15 min

State assumptions, derive cleanly, check dimensions or limiting cases, then interpret the result.

  • Setup & assumptions
  • Derivation
  • Sanity checks
  • Interpretation
  • Communication

The attempt–diagnose–repair loop

  1. Attempt closed-book. Use the round timer and speak, derive, design, or code as the real format requires.
  2. Mark observable weak dimensions. “Felt bad” is not a diagnosis; “never defined the ship decision” is.
  3. Select overall evidence. Weak is due in two days; Review in one week; Confident counts toward (but does not complete) graduation.
  4. Read only to repair. Compare with the reference after scoring. Write the missing mechanism, decision, test, or story evidence.
  5. Retry after spacing. Use a changed follow-up or prompt. Immediate fluency after reading is not retained evidence.
  6. Interleave later. Verify the repair when the category is no longer predictable.

Graduation rule

For a critical question or pattern, require two successful closed-book attempts on different days, no critical rubric dimension marked weak, one unfamiliar follow-up, and survival in a later mixed session.

How to record yourself

  1. Record audio or screen locally; do not upload confidential project detail.
  2. Watch once for content: correctness, decisions, evidence, and omissions.
  3. Watch again for delivery: answer fit, structure, hedging, silence, and response to pushback.
  4. Write one corrective behavior. Do not produce a transcript of every imperfection.
  5. For L5/L6, ask a peer to challenge at least two assumptions rather than simply listen.

A standard 60-minute short-answer block

08 min

Closed-book answer

07 min

Rubric diagnosis

20 min

Targeted repair

08 min

Rewrite outline from memory

12 min

One changed follow-up

05 min

Schedule retry

Self-score limitation

A local score cannot calibrate L6 scope, story ownership, implementation correctness, or room presence by itself. Use at least one observed simulation and instruct the observer to push back.

Choose a question → · Open the retry queue →