🗺️ Navigating the loop¶
When: the day before. What each round is actually scoring.
Companies weight these differently, but the rounds themselves are recognisable across most retrieval and applied-ML loops.
The rounds¶
| Round | What it looks like | What is being scored | The way people lose it |
|---|---|---|---|
| Screen | "Walk me through a project" | Can you describe your own work at the right altitude? | Narrating the architecture instead of the decision. They want the fork in the road |
| Technical depth | "Why does X work?" | Do you know the mechanism or the name? | Reciting the name of the technique confidently. Every follow-up goes one level down until you stop |
| System design | "Design retrieval for…" | Do you find the constraint that makes it hard? | Designing the finished system and never saying what week one is |
| Case / take-home | Data and a vague goal | Do you measure before you tune? | Shipping a model. They wanted an eval set |
| Behavioural | "Tell me about a disagreement" | Do you change your mind on evidence? | Stories where you were right the whole time |
The three things scored in every round¶
1 · Do you reach for evidence or intuition? "I'd check the trace" beats "I'd look at whether the right documents came back." The first is a procedure, the second is a wish.
2 · Do you know what you cannot see? Naming a missing trace, a missing eval set, or a missing label as a finding — rather than as a blocker — is the single clearest seniority signal available to you.
3 · Have you actually done it?
One concrete failure you lived through beats any amount of taxonomy. "A team I worked with spent
a month on the encoder for a problem that was k=5" does more work than a correct list.
The first five minutes of a design round are scored¶
Most candidates start designing. The clarifying questions are the round.
Ask, in this order:
- What is the constraint that makes this hard? — cost, latency, permissions, freshness, residency. There is always one; find it before you draw anything
- How would we know it works? — if they have no answer, that is the first deliverable
- What is the cost of a wrong answer versus a refusal? — decides abstention, and almost nobody asks
Two sentences that reliably score¶
"Before I design this — can a user know a document exists but not read it? That changes the architecture, not just a detail of it."
And:
"Week one is not any of this. Week one is an eval set and a test that fails when we break it, because otherwise we cannot tell in March whether March was better."
Managing the 45 minutes¶
| Minutes | Do |
|---|---|
| 0–5 | Clarify. Write the constraint on the board. Do not start designing |
| 5–20 | Design out loud. Say the alternative you rejected and why |
| 20–35 | They will pressure one weak point. They have already chosen it — it is usually the thing you glossed |
| 35–42 | "What would you do first?" Answer with week one, not the architecture |
| 42–45 | Your questions. Ask one real one |
When you do not know¶
Say so, then say what you would do. "I don't know how HNSW handles deletes specifically. I'd expect a tombstone-and-compact scheme because rebuilding the graph per delete is too expensive, and I'd verify that before designing around it."
That scores better than a confident wrong answer, and better than "I don't know" alone. The reasoning is the thing being measured.