Capstone: Mock Loop

Set a 12-minute timer and answer each prompt out loud, as if a RevOps lead and a controller are both in the room. The technical screen is part interview, part oral exam: they will follow up on whatever you say, so the goal is to teach each topic cleanly, not recite a definition.

Cursor sells a consumption-heavy product, so the rapid-fire bends toward usage. If your ASC 606The US revenue-recognition standard; cited as the canonical judgment-heavy accounting work to keep human-led rather than hand to an agent, because facts and circumstances vary deal to deal. answer works for a flat annual subscription but falls apart the moment metered usage enters, you have not yet earned the room.

1. The five steps of ASC 606 and where usage breaks eachAnswer in under 90 seconds

2. Draw Order-to-Cash and name the system of record at each stageSystem of record is the real test

Anyone can list the stages. The signal is naming which system owns the truth at each hop, because that is what you will spend a year integrating.

1. 1Lead / opportunity. CRM (Salesforce-class) owns the account and pipeline.
2. 2Quote / CPQ. CPQ owns the configured price, ramp and discount approvals; it emits the order.
3. 3Order. The billing platform (Stripe / Metronome / Orb / Zuora-class) owns the subscription, plan and commercial terms.
4. 4Provisioning / entitlement. The product or an entitlement service owns what the customer is allowed to consume.
5. 5Usage metering. A metering pipeline owns the raw consumption events and the rated quantities.
6. 6Invoice. The billing platform owns the invoice as the legal billing document.
7. 7Payment / cash application. The payment processor captures cash; the ERP (NetSuite / Oracle / SAP-class) owns cash application against the AR subledger.
8. 8Dunning / collections. Billing or a dedicated dunning tool owns retries and the collections workflow; the ERP owns the AR aging.

3. Where the revenue subledger sits and why the GL alone is not enoughThis separates engineers from configurators

A general ledger stores balances: it knows you have $4.2M of deferred revenue this month. It does not know which contract, which performance obligation and which schedule produced that number. Auditors and your own restatement-prevention instincts need that lineage.

4. Keeping usage, invoices and the GL reconciled at scaleThree-way tie-out

• Usage → invoiced. For each billing period, the sum of rated usage events must equal the usage charges on the invoice. Reconcile on a stable usage key with a tolerance threshold and alert on any drift.
• Invoiced → recognized. Recognized revenue ties to the rev subledger schedules; deferred plus recognized must equal billed plus any unbilled receivable.
• Subledger → GL. The subledger's posted journals must foot to the GL revenue and deferred-revenue accounts every period, with a documented reconciliation owner.
• Controls. Immutable usage ledger, idempotent ingestion and a daily three-way reconciliation job that fails loudly rather than silently absorbing variance.

5. A build-vs-configure decision with real trade-offsHave one rehearsed

Open a blank query editor and a 25-minute timer. The technical screen often lets you use Cursor or another AI assistant for targeted lookups, but the model answer is yours to defend. Narrate trade-offs as you type, the way you would with someone watching your screen.

Drill A - deferred-revenue roll-forwardThe query a controller asks for monthly

A roll-forward proves the deferred balance moved correctly: opening + additions (billings) − recognized = closing. If it does not foot, you have a recognition bug.

WITH movements AS (
  SELECT
    date_trunc('month', event_date) AS period,
    SUM(CASE WHEN type = 'billing'     THEN amount ELSE 0 END) AS additions,
    SUM(CASE WHEN type = 'recognition' THEN amount ELSE 0 END) AS recognized
  FROM rev_subledger_events
  GROUP BY 1
)
SELECT
  period,
  SUM(additions - recognized)
    OVER (ORDER BY period
          ROWS BETWEEN UNBOUNDED PRECEDING AND 1 PRECEDING) AS opening_deferred,
  additions,
  recognized,
  SUM(additions - recognized)
    OVER (ORDER BY period
          ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW) AS closing_deferred
FROM movements
ORDER BY period;

Drill B - usage → invoiced → recognized reconciliationThe three-way tie-out as one query

WITH usage AS (
  SELECT contract_id, billing_period,
         SUM(rated_amount) AS usage_rated
  FROM usage_events_rated
  GROUP BY 1, 2
),
invoiced AS (
  SELECT contract_id, billing_period,
         SUM(line_amount) AS invoiced_amount
  FROM invoice_lines
  WHERE charge_type = 'usage'
  GROUP BY 1, 2
),
recognized AS (
  SELECT contract_id, billing_period,
         SUM(amount) AS recognized_amount
  FROM rev_subledger_events
  WHERE type = 'recognition'
  GROUP BY 1, 2
)
SELECT
  COALESCE(u.contract_id, i.contract_id, r.contract_id) AS contract_id,
  COALESCE(u.billing_period, i.billing_period, r.billing_period) AS billing_period,
  u.usage_rated, i.invoiced_amount, r.recognized_amount,
  u.usage_rated   - i.invoiced_amount   AS usage_vs_invoiced_diff,
  i.invoiced_amount - r.recognized_amount AS invoiced_vs_recognized_diff
FROM usage u
FULL OUTER JOIN invoiced i USING (contract_id, billing_period)
FULL OUTER JOIN recognized r USING (contract_id, billing_period)
WHERE ABS(COALESCE(u.usage_rated,0) - COALESCE(i.invoiced_amount,0)) > 0.01
   OR ABS(COALESCE(i.invoiced_amount,0) - COALESCE(r.recognized_amount,0)) > 0.01;

Drill C - the schemaContracts, entitlements, usage, invoices, cash

Drill D - idempotency and late events on paperMetering pipeline hygiene

1. 1Idempotent ingestion. Give every usage event a producer-generated stable event_id; INSERT with ON CONFLICT (event_id) DO NOTHING so a retried delivery cannot double-count.
2. 2Event-time vs. ingest-time. Store both event_time and ingested_at. Rating and recognition key off event_time so you assign usage to the period it actually happened, not the period it arrived.
3. 3Late-arriving events. A late event reopens the affected billing period: re-rate that period, post a delta to the subledger and emit a true-up line on the next invoice rather than mutating a closed, audited invoice.
4. 4Watermark and cutoff. Define a close cutoff (for example T+2 days) after which a period is locked; events past the watermark route to a true-up, with the variance tracked as a control metric.

Drill E - three data-quality controls and their queriesEach control needs an enforcing query

This is the round the role is really hiring for. Set a real timebox (90 minutes for practice, the day for the real thing), open Cursor and treat the prompt as deliberately under-specified. They are grading your judgment under ambiguity, not a finished artifact.

The promptRead it twice, then scope it

Design and partially build a usage-to-invoice pipeline (or a reconciliation / anomaly service) for a consumption-priced AI product. You have access to AI tools. Show us how you think. That is roughly the whole brief - the rest is yours to define.

Minutes 0–15: scope ruthlessly and state assumptionsEarn the right to build

1. 1Write the assumptions down first. One metric (tokens), one plan shape (fixed rate per 1k tokens with a monthly minimum), one currency, monthly billing. State them as choices you would confirm with RevOps, not as facts.
2. 2Pick the thin vertical slice. Raw usage event → idempotent landing → rated → aggregated to an invoice line, with one reconciliation check. Explicitly defer multi-currency, tax, dunning and CPQ.
3. 3Name the proof. Decide up front what 'done' demonstrates: re-running ingestion does not double-count and the usage total ties to the invoiced total.
4. 4Say the meta-point. “I'm optimizing for proving the system design end-to-end over feature coverage, because a half-built sprawl tells you less than one slice that runs and ties out.”

Minutes 15–70: build the slice, narrate the AIThe AI-authenticity test is live here

What the slice must coverHit these four even if you cut features

On the agent boundary, be specific to finance. An agent drafting a reconciliation summary or flagging a 30% usage spike for human review is safe. An agent posting a journal entry to the GL without a human approval and an immutable log is not, because it breaks segregation of duties and your audit story.

Minutes 70–90: the closeoutHow you end is how they remember you

1. 1What's built. Demo the proof live: run ingestion twice, show the count unchanged, show the tie-out passing.
2. 2What's next. The deferred list, in priority order: tax, multi-currency, prepaid credits, dunning - and why each was safe to defer for this slice.
3. 3Risks I deliberately deferred. Name them as risks, with the control you would add: 'no SOXSarbanes-Oxley Act. A US law that forces companies to keep auditable controls over any system that affects their financial reporting. evidence yet; I'd add an approval gate and an immutable audit log before this touches a real GL.'
4. 4The one trade-off you'd revisit. Pick the assumption you are least sure about and say what would change it.

Record yourself on your phone and play it back. The behavioral and founder rounds screen for intensity, ownership, autonomy and genuine conviction - and those land through evidence and specifics, not adjectives. If your story would survive a skeptical follow-up, it is ready.

The three spine stories, each under three minutesTime them; cut the setup

Pressure-test 'why Cursor' for specificityWould it survive a follow-up?

A generic 'I love AI and dev tools' answer dies on the first follow-up. Anchor it to facts you can defend: this is Cursor's first finance systems hire, the product is consumption-priced and the company scaled to multi-billion ARR fast, which makes usage-based billing and its ASC 606The US revenue-recognition standard; cited as the canonical judgment-heavy accounting work to keep human-led rather than hand to an agent, because facts and circumstances vary deal to deal. treatment the most relevant domain you could be working on.

Translate a vague commercial requirement, liveThe cross-functional round in miniature

They will hand you something fuzzy from Sales or Legal and watch you turn it into a technical plan without losing the commercial intent. Practice the loop until it is reflexive.

1. 1Restate the intent. “What Sales wants is to offer a 12-month ramp deal with usage credits that roll over - let me make sure I've got that.”
2. 2Surface the hidden questions. Do credits expire, are they refundable, do they apply before or after the monthly minimum, how does a mid-term upsell interact with the rollover.
3. 3Name the system impact. Entitlement model, rating logic and the ASC 606The US revenue-recognition standard; cited as the canonical judgment-heavy accounting work to keep human-led rather than hand to an agent, because facts and circumstances vary deal to deal. question of whether rollover credits change when revenue is recognized.
4. 4Propose the cut. What you'd build first to unblock the deal and what you'd defer with a clear flag back to Finance.

Have your senior-signal questions readyQuestions are part of the evaluation

• “What's the current state of the finance stack - what's configured, what's duct-taped and what's the first fire?”
• “Where does revenue recognition live today and who owns the close - is there a subledger or is it spreadsheet-driven right now?”
• “How do you want to balance shipping fast against SOXSarbanes-Oxley Act. A US law that forces companies to keep auditable controls over any system that affects their financial reporting.-readiness as you approach IPO scale?”
• “What does the first 90 days look like and how will we know it went well?”
• “Where have AI agents already been let near financial data and what guardrails exist?”

Score each row honestly on a 1–5: 5 means you could teach it cold, 3 means you can recall it but would wobble under a follow-up, 1 means you'd freeze. Anything below 4 is a study target. The bar for this role is teaching, not recognition.

Treat the rubric as a gap finder, not a score to admire. Mark the weakest proof, turn it into one practice rep and keep the artifact you would show an interviewer.

Domain readinessCan you teach it, not just recall it

Technical readinessUnder time, narrating trade-offs

Build & behavioral readinessThe two rounds with the highest variance

Logistics - the part candidates forgetConfirm before the loop, not during

• Confirm the per-stage AI policy with your recruiter. Which rounds permit Cursor and AI tools and for what - targeted queries only or full build access. Do not guess on the day.
• Be fluent in Cursor live. If you'll build in it, rehearse until inline edits, agent runs and accepting/rejecting diffs are reflexive - a fumbled tool is a self-inflicted wound in an AI-native loop.
• Have your environment ready. A scratch repo, a SQL sandbox and your spine stories in a doc you won't actually read from.