You are in Round 1.

The interviewer is Raj — Staff Engineer, ads platform team at StreamCo. He has been in the chair on your side of this conversation four times in the past two weeks. He has seen enough candidates freeze on this round to know that the freeze is the thing being scored.

Raj nods once. Timer reads 4:31. He says:

Four numbers, one word.

Raj is watching. Forty minutes of round left. He asked for the reconciliation. You produce it out loud, in SQL, while naming the four revenues StreamCo actually runs on — because at this scale they do not collapse into one.

You begin. Speaking to Raj, keyboard in your lap, you say:

Raj nods. You draw four boxes on the sketch pad you have shared to your screen.

You turn back to Raj and continue:

Raj pulls up the shared editor on his side. You start typing.

The query finishes at 10:21. You screenshot the result. Raj watches you do it without comment. You then write the paragraph the director has been waiting for.

You close the editor. You look up at Raj. The timer reads 21:14.

Raj reads the paragraph twice. He says:

He changes his screen back to the whiteboard you had drawn 17 minutes ago. The half-finished architecture is still there — Kafka at the left, a few scattered boxes, the words attribution underlined twice. He looks at it. He looks back at you.

The whiteboard is yours. The incident is your brief. The clock keeps running.

Build a platform that makes this morning impossible.

Clock reads 21:14. Raj wipes the whiteboard. He does not say a word. He caps the marker, sets it down where you can reach it, and leans back. This is the part of the round where most candidates start drawing boxes and arrows. This morning is fresh in your head. You do not reach for the marker yet.

Before you pick up the marker you say one sentence. It is the sentence that frames every box you are about to draw. You want this sentence in Raj's head before a single rectangle appears on the board, because every architectural decision you make after this is a consequence of it.

Now you draw. Not a wiring diagram — a layer stack. Five layers, top to bottom. Each layer does one thing. Each layer has a contract with the layer above and below. You draw them as stacked rectangles.

The stack is on the board. Raj has not interrupted. You have four minutes of air, so you spend them on the three contracts that make the stack real. Without them the boxes are furniture.

The SQL shape of the append-only attribution table is the single hardest thing to get right, so you write it on the board:

-- gold.revenue_attributed · append-only, bi-temporal
CREATE TABLE gold.revenue_attributed (
  conversion_id         UUID        NOT NULL,
  advertiser_id         BIGINT      NOT NULL,
  campaign_id           BIGINT      NOT NULL,
  event_time            TIMESTAMP   NOT NULL,  -- when the conversion happened
  attribution_time      TIMESTAMP   NOT NULL,  -- when we credited it
  attribution_model     VARCHAR(32) NOT NULL,  -- last_click_7d, mta_linear, ...
  identity_graph_version VARCHAR(16) NOT NULL, -- pinned for reproducibility
  credited_usd          NUMERIC(18,4) NOT NULL,
  is_superseded         BOOLEAN     NOT NULL DEFAULT FALSE,
  PRIMARY KEY (conversion_id, attribution_time)
)
PARTITION BY RANGE (attribution_time);

-- "what did we show Finance on the 10:37 email on 2025-09-30?"
SELECT SUM(credited_usd)
FROM   gold.revenue_attributed
WHERE  event_time >= '2025-09-01' AND event_time < '2025-10-01'
  AND  attribution_time <= '2025-09-30 10:37:00'
  AND  is_superseded = FALSE;

That last query is the one that makes the platform trustworthy. Any number Finance ever saw is reproducible by plugging the attribution_time ceiling into the same query. "What did we tell you, and when?" becomes a SQL answer instead of a Slack archaeology project.

Raj picks up the marker. Not to draw. To point.

Every box gets punched. You defend or you fold.

Clock reads 24:03. Raj has the marker. He is not drawing — he is pointing. He will spend the next nineteen minutes trying to break the architecture you just put on the board. Each challenge is a real bet he has watched candidates lose. You have one correct shape of answer for every question: name the alternative, name the trade, name the harm the trade prevents.

There is no clean way to script this. So here's how it actually goes: six volleys, a clock, and a running counter of whether you are converting pressure into signal or leaking it.

Clock reads 43:22. Raj caps the marker. The forty-five-minute round ends in eighty-eight seconds. You think you are done. You are not.

Black Friday. 2:17 AM. ROAS up 18%.

The platform shipped in late October. The CFO's Friday dashboard has run clean for ten weeks. You are asleep. PagerDuty does not care.

The guardrail is your own. You wrote it in week six, after Raj's 2:17 AM question in the interview turned out to be a real hour at this job. The alert fires on WoW ROAS delta >4% across the top-five advertisers. Eighteen percent is not a drift. Eighteen percent is a bug, a miracle, or a fraud. Finance sees this number at 09:00. You have roughly thirty minutes before the director wakes up and sixty before that meeting starts.

The thing about debugging your own platform is that you know exactly where to look. The architecture is not a mystery. The bug will be in exactly one of four places, and the platform was designed to tell you which one.

H3 is hot. You write the query your architecture made cheap:

-- ROAS decomposed by identity graph version, last 24h
SELECT
  identity_graph_version,
  COUNT(DISTINCT conversion_id)                 AS n_conversions,
  SUM(credited_usd)                             AS attributed_usd,
  SUM(credited_usd) / NULLIF(s.spend_usd, 0)    AS roas
FROM   gold.revenue_attributed r
JOIN   gold.revenue_spend      s USING (advertiser_id)
WHERE  r.advertiser_id IN (/* top 5 */)
  AND  r.attribution_time >= NOW() - INTERVAL '24 hours'
  AND  s.spend_date = CURRENT_DATE
GROUP BY identity_graph_version, s.spend_usd
ORDER BY identity_graph_version;

You do not revert. You do not go back to bed. You write the paragraph.

03:21 AM. You close the laptop. You lie down. You stare at the ceiling for forty minutes. You are still awake at 4:02 when the 4:00 AM batch runs green. You already know the thing the last hour actually tested, and it is not the thing Raj asked.

08:29 AM. The CFO's office. Two dashboards, one conversation.

Five hours after the page, you walk into Amelia's office with a laptop and two browser tabs open. One is the dashboard the platform ships to her every morning. The other is a parallel view you built at 03:40 AM to explain tonight. The meeting is seventeen minutes. You do not open your laptop for the first three.

The meeting is not about the platform. It is about whether the platform can narrate itself. If the dashboard cannot answer "why is tonight's number what it is" without you in the room, the platform is not done, no matter how many tests pass. So you open the laptop. You start with the picture she sees every morning.

The CFO's morning view · conversion funnel

This is what Amelia sees at 07:45 AM, every day, on her phone. One chart, top of the dashboard. Five steps. It answers: how did the whole ad tier perform yesterday, shape-wise. It does not answer a question about dollars — that's the second section below. This chart answers whether the shape of the day was normal.

The funnel · from impression to repurchase

The first thing you point at on the screen is the drop-off between two specific steps. Not the overall number. The shape. Shape reveals measurement changes before totals do.

The tab control at the top is the feature that matters this morning. Switching attribution window from 7-day to 1-day to 28-day shifts the tail of the funnel. Last night's identity-graph change affects 28-day attribution most — by design, more cross-device journeys close within 28 days than within 1. Shape tells the story without dollars.

ROAS by attribution window · the reason you're here

This is the chart you built overnight. It is not in Amelia's normal view. You pulled it forward because the only way to explain tonight's +18% is to show ROAS across attribution windows side by side, with the identity-graph version as a visible cohort.

You say one sentence while it loads: "The 28-day window is where the graph change shows up. 7-day barely moves. 1-day is flat. The platform is measuring more accurately — not more optimistically." Amelia looks at the 28-day bar for thirty seconds. She asks one question. It is the question every CFO asks eventually, and the platform has to have a clean answer.

You were ready for this one at 03:40 AM. You built the parallel table for exactly this question.

She reads the email for ninety seconds. She does not say anything while she reads. When she looks up, the first thing she says is not about the numbers.

You answer honestly.

She nods once. The meeting ends at 08:42. The board prep gets the $9.31M footnote. The advertisers get the note on Tuesday. Nobody mentions the word "incident."

The advertiser view · install-to-event funnel

After the CFO meeting you walk down to the advertiser-success team. They have their own dashboard — the one the platform ships to every self-serve advertiser. This is a different surface for a different audience, and the question it answers is different too. Is my campaign working, in my vocabulary.

The difference between this view and the CFO view is the thing that took you three months to learn at StreamCo: audiences do not translate. Amelia does not want to see "D30 retention by cohort." An advertiser does not want to see "settled revenue, USD basis." The same ad platform produces both surfaces from the same silver tables, bound through the semantic layer, and neither audience is ever confused about which number is which — because the audience tag on the metric is part of the query plan.

The sales-cycle view · lead funnel + cycle time

The third surface is for advertisers running lead-generation campaigns. The shape of a lead campaign is not a funnel; it is a funnel with a time axis, because the bottleneck is days-to-close, not percent-converting. A lead campaign with a 40% MQL-to-close that takes 90 days is worse than a 25% MQL-to-close that takes 14 days, and no one-number ROAS captures this.

This dashboard is the one advertisers screenshot and send to their CFOs. The platform's job is to make sure they don't have to re-label anything before they paste it into the deck — the axis labels, the window, the attribution model are all on the chart itself.

There is one more dashboard below — the campaign-level drilldown. It is a tool, not a narrative surface. You include it because every advertiser eventually asks "but what about my campaign," and the platform needs a no-ceremony way to answer that.

Campaign drilldown · the header card

Single campaign, scorecard at the top, trend below, benchmark row for context. Self-serve advertisers bounce between this and the funnel chart above; that pattern drove the layout decisions.

The last four minutes.

Clock reads 43:22. Raj caps the marker for the second time. The 2:17 AM story is on the whiteboard. The CFO meeting is on the whiteboard. There are no more scenarios. You think the round is done. Raj leans forward, not back, which is how you know it is not.

You do not answer for six seconds. This is not thinking time. This is the moment the round pivots, and you need the pivot in your own head before you put it into his.

You say it as one sentence because the question asked for one.

Raj does not nod. He also does not write anything down. The absence is the signal. He reaches for the laptop on the side of the desk and closes it — not yours, his. The round is effectively over. He spends the remaining thirty seconds on one thing.

He lets the last principle sit for eight seconds. Then he stands up, which is the exit move he has made on every good candidate in the last year. The interview ends in an unremarkable sentence.

You do not ask what aligned means. You already know. The round scored you on the only thing that matters at this level: whether you know the enemy in your own system. You named it in Act 1 — silent label drift between audiences who think they are reading the same number — and every act after that was one more test of whether you'd forget it under pressure.

You did not forget it. The offer comes three days later.