2 min lesson
Quantify in the units this charter respects
Compare two rows from "Quantify in the units this charter respects", then say when each one fits.
Step 1 of 2
Quantify in the units this charter respects
- Weak result
- “Improved GPU efficiency”
- TPM-grade result
- “Lifted sustained utilization from ~22% to ~48% on the shared training pool, ~$2M/yr”
- Weak result
- “Helped with cost attribution”
- TPM-grade result
- “Stood up cost-per-active-user tracking that resolved 90% of spend to a product within one quarter”
- Weak result
- “Ran the capacity planning”
- TPM-grade result
- “Built the demand model that tied GPU need to DAU growth; the next-quarter forecast held within 8%”
- Weak result
- “Drove the migration”
- TPM-grade result
- “Sequenced the inference-cluster migration across 4 teams with zero SLA breach, on the committed date”
| Weak result | TPM-grade result |
|---|---|
| “Improved GPU efficiency” | “Lifted sustained utilization from ~22% to ~48% on the shared training pool, ~$2M/yr” |
| “Helped with cost attribution” | “Stood up cost-per-active-user tracking that resolved 90% of spend to a product within one quarter” |
| “Ran the capacity planning” | “Built the demand model that tied GPU need to DAU growth; the next-quarter forecast held within 8%” |
| “Drove the migration” | “Sequenced the inference-cluster migration across 4 teams with zero SLA breach, on the committed date” |
Replace adjectives with a number and the unit ($/yr, utilization %, forecast error %, SLA breaches).
Prepare two stories that mirror the charter exactly. One “I influenced senior technical leaders to commit to a plan I couldn't mandate” story - the influence-without-authority signal. One “I built the model myself and the model changed the decision” story - the action-bias-plus-altitude signal. If you only prep one, prep the influence one, because that's the whole job.
“We cut idle GPU spend about 22% in a quarter, roughly $3M annualized. The hard part wasn't the analysis - it was that three teams shared the pool, none of them reported to me and each assumed the others were the wasteful one. So I built the attribution model that showed where the idle hours actually sat, brought it to a weekly review I ran and let the data settle the argument instead of me. The ML leads owned the model-routing change that did most of the work; I owned getting them to agree on the split and holding everyone to it. If I redid it, I'd have instrumented per-team utilization a month earlier - we argued from anecdotes for too long before the data existed.”
Learn more
Full explanation
Full explanation
Two things sink a TPM story. An all-“we” narrative where the interviewer can't find your specific contribution. And a failure story that's a humblebrag (“my weakness is I care too much about accuracy”). Keep one genuine failure where a program slipped or a forecast was wrong and land it on the real lesson, not a disguised strength.
QYou're telling a GPU cost-savings story. Which Result sentence is strongest for this role?