Skip to content
mlmentorship

Tell me about an ML project you killed or redirected

A staff-level behavioral and project question about sunk costs, evidence thresholds, influence, and responsible stopping.

Published · 3 min read ·Core ·Mixed

Behavioral / leadership · active recall

Practice before you read

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

Back to my workbook How practice works

Behavioral / leadership · closed-book attempt

Tell me about an ML project you killed or redirected

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

03:00recommended time

Closing or reloading clears the scratchpad. Only score, weak rubric dimensions, attempt count, and retry date can be stored locally.

30-second answer map

Visual first · depth when needed

Decide whether to fund one more bounded experiment or stop and redirect by comparing only forward evidence, opportunity cost, and transition value rather than past spend.

Preparing the visual…

Asked in: staff-level behavioral and project rounds.

The L4 story is “the model missed its accuracy bar so the team cancelled it.” The L6 story is a project you stopped on evidence, against sunk cost and your own attachment, then redirected the people and infrastructure to something higher-leverage. Killing the right project on time is a senior signal; treating only launches as impact is a junior one.

What a strong answer covers

Explain the decision, not only the outcome. Cover:

  1. Why the project was rational at the start.
  2. Which evidence or failed assumption changed the case.
  3. Why another iteration was a poor use of resources.
  4. What experiment, proposal, and transition work you owned.
  5. The cost avoided, effort redirected, users protected, or capability learned.
  6. The earlier signal you would instrument next time.

Learning objective

What belongs in the decision when another iteration is competing with a redirect?

FIRST, CLOSE THE SUNK-COST COLUMN

Past spend explains the stakes; it does not earn another quarter.
Record what was invested and learned, then compare only what each option costs and returns from today forward.

FUND ONE BOUNDED TEST

Condition
A named uncertainty could reverse the recommendation, and a cheap decisive test can separate a broken thesis from fixable execution.

Commitment
Pre-agree the success bar, budget, owner, and stop date. One test buys information, not automatic permission for another loop.

STOP AND REDIRECT

Condition
The core assumption has failed, or expected future value is lower than the next-best use of the same people and infrastructure.

Transition
Protect users, move ownership deliberately, and preserve reusable data, tooling, infrastructure, and lessons for the better path.

Read it this way: read past investment once, as context, then close that column. Continue only when a bounded experiment could change the decision; otherwise compare the project's future return with the opportunity cost and redirect without discarding reusable work. This original decision ledger was checked against Arkes and Blumer's sunk-cost experiments, NASA's decision-analysis guidance, and OpenStax's treatment of opportunity cost.

What an L4 answer sounds like

“The model did not reach accuracy, so the team cancelled the project.”

An outcome with no judgment or ownership in it.

What an L5 answer adds

An L5 answer owns the evidence and recommendation inside a project area. It defines minimum viable quality before more spend, runs an experiment that separates a broken thesis from fixable execution, identifies reusable work, and transitions users, data, and infrastructure responsibly.

What an L6 answer adds

An L6 answer manages the portfolio and organizational consequences: pushing past sunk-cost and prestige incentives, aligning several teams on a new direction, keeping the trust of sponsors who backed the original bet, and changing review criteria so similar weak bets fail faster. Keep the technical detail. “I reallocated the roadmap” with no evidence behind it sounds managerial and vague.

Tells that get you a strong-hire vote

  • The original decision was reasonable, not an obvious straw man.
  • You name the assumption that failed.
  • The stopping criteria were evidence-based.
  • You handled the people and transition costs with care.
  • You account for both avoided waste and lost opportunity.

Tells that get you down-leveled

  • The project died because another team cancelled it, not because of you.
  • You take credit for hindsight.
  • No one who disagreed is represented fairly.
  • “We pivoted” hides the specific decision.
  • Nothing you learned changed how you invest next time.

Common follow-ups

  • Why not iterate for another quarter?
  • Who disagreed, and what evidence did they have?
  • How much had already been invested?
  • What parts of the work were reusable?
  • Tell me about a time you stopped too early, or held on too long.

Related: decide what to work on, scope an ambiguous problem, and senior through senior-principal ML scope.

Found an error or stale claim? Report a correction or send a private note.