Present a technical ML project through its problem, your decisions, supporting evidence, and outcomes while handling detailed follow-up questions.
Problem
Context. An interviewer requests a project you personally worked on; the fictional critical-path trace below is a separate presentation rehearsal.
Task. Explain the goal, ownership, constraints, mechanism, alternatives, validation, and outcome of your project. For the rehearsal, choose the first dependency to investigate and state what one trace cannot establish.
Constraints. Distinguish your contribution from the team’s and observed results from estimates or counterfactuals. Do not claim the fictional example as personal experience or infer a population p99 from one request trace.
In this fictional rehearsal, the complete request has a 100 ms p99 objective. One critical-path trace starts at 0 ms:
| Parallel dependency | Completion |
|---|---|
| Model inference | 45 ms |
| Feature join | 120 ms |
The response needs both results; other stages are omitted from this simplified trace. Which component should you investigate first? What does this trace not establish?
Picture: A Question Opens One Decision Record
Decision records
A faster model cannot bypass the slower dependency that the response awaits
Derivation
Chronology alone
Setup → data → model → launch
This order can hide ownership and alternatives. A question may require reconstructing an earlier decision.
Independent records
Decision 1 · Decision 2 · Decision 3
Each carries constraint, alternatives, prior evidence, choice/owner, result and update. Open the requested record directly.
One request, two parallel dependencies
Which dependency remains on the end-to-end critical path?
Investigate the Blocker Without Overclaiming
Completion cannot precede ms in this trace. Making inference instantaneous would still leave the join at 120 ms.
Investigate a versioned join change under matched load and inspect complete request latency and quality. A single trace does not estimate p99; stage percentiles must not be added into an end-to-end claim. State your actual ownership and other teams’ changes.
The Decision Spine
Use this editable record for each real consequential choice:
| Field | Record |
|---|---|
| Constraint | What blocked the obvious approach? |
| Alternatives | Which credible options existed? |
| Evidence then | What was known before the outcome? |
| Choice/owner | What did you decide, and with whose authority? |
| Result | What happened, including limits or actual surprises? |
| Update | What changed, or would change under new evidence? |
Connect one record to a sanitized calculation, trace, ablation, data contract or invariant. Do not invent failure, metrics, ownership or a causal interpretation to fill fields.
Change 1: Ten Minutes Remain
New task. Time shrinks or questions interrupt early.
Picture and change. Keep the opening claim and one complete record; place other evidence in appendices.
Check. Remove repetition, not the reason for the choice. Answer directly before returning to the deck.
Use when. Delivery time or question order changes.
Change 2: Resources or Criticism Change
New task. Compute is halved or a credible critic disputes the reported effect.
Picture and change. Re-evaluate constraints and evidence; identify which intervention to retain and which claim changes.
Check. Do not defend historical architecture after its assumptions fail.
Use when. A changed condition tests your model of the system.
Failure: Impressive Metrics Without an Owned Choice
An architecture tour cannot substitute for a credible alternative, owner, supporting mechanism and evidence limit.
Transfer Check
Why does a holdback alone not prove your impact claim?
Reasoned Answer
Causal interpretation depends on assignment, exposure, measurement, interference, attrition and analysis assumptions. Name the actual design and limits.
Choose the project with the best evidence
Check five properties: owned consequential judgment, credible alternatives, actual uncertainty/surprises, outcome evidence or explicit limits, and implementation depth. Prestige, duration and a compulsory failure story are not substitutes. Respect confidentiality and approved disclosure boundaries.
Open with the answer
State problem, importance, personal responsibility, observed result/limits and the decision roadmap. A two-minute opening, three records and 30-minute talk plus 15-minute questions are configurable rehearsal settings, not universal interview facts.
Evidence Ledger
| Claim | Source/version | Personal action | Collaborator | Observed outcome | Causal design/limit | Changed condition |
|---|---|---|---|---|---|---|
| Fill with approved real evidence | Exact artifact/trace | Your contribution | Other owners | Measured or unavailable | Identification assumptions | Reopening evidence |
Use the same names at overview, architecture and implementation depth. Explain adoption, incentives and second-order effects only when supported; distinguish your contribution from the project’s total result.
Check the Account
Can a question enter through evidence, ownership, mechanism or changed constraints without losing the claim? If evidence is unavailable or confidential, state the limit rather than inventing a number.
Missing Evidence
Replace vague team attribution, offline-as-product claims and unsupported strategic language with the actual decision record.
Common follow-up
What would your strongest critic say was wrong? State the actual criticism, supporting evidence, agreement/disagreement and decision to revisit. Do not manufacture a critic or substitute a disguised compliment.
Use the presentation protocol and attempt it before reading this page again.
Sources
Related: your most ambitious project, scope an ambiguous problem, and technical presentation practice.