You have worked the five modules — the foundation, the query & join engine, quality & root cause, experiments & charts, and the reporting system. Each one ended with a graded artifact: a profiled dataset and a metric dictionary with named owners and edge cases; a query a stranger could audit and a joined dataset reconciled to the row and the dirham; a quality scorecard with stated rules and severities, and a root-cause writeup that ruled out an artifact before it confirmed a real billing bug; an experiment readout that survived a skeptic and a chart verified before it was rendered; a recurring monthly report and a self-serve analytics system a non-analyst can trust. The capstone is where those five stop being five.
This is the integrative project that earns the credential. One brand, one analytics function, end to end — a complete system where the metric dictionary is still load-bearing in the recurring report, the queries trace to the same definitions the quality audit tested, and the experiment readout cites the same metrics.md the monthly report does. That is the thing “Certified Data & Analysts with Claude” attests, and it is a higher bar than five modules completed: a certified analyst is not someone who watched the track, it is someone who can run a complete analytics function where every number traces to a source and every figure that leaves the folder has already survived the second question.
Five modules teach you the moves. The capstone proves you can run them as one system — a complete analytics function where the profile, the dictionary, the queries, the quality gate, the experiment readouts, and the recurring report all inherit the same source of truth. A reviewer should not be able to tell that the metric dictionary, the quality scorecard, and the May board-pack handoff were drafted in different sessions for different problems. Integration is the grade, not five separate artifacts that each look fine alone.
The brief
Produce a complete analytics function for one brand, and prove it by running that team’s data operation from the raw export to the recurring report, end to end. We recommend your own data — the capstone then doubles as deployable infrastructure you keep and run (see Bring your own data) — or the sample, Mizan’s analytics function, the GCC bookkeeping SaaS whose Business Analyst, Maya Haddad, is threaded through the whole track, if you’d rather learn on neutral ground.
You already know the Desktop motion: open the team’s folder in Claude Desktop, approve each read in the “Ask permissions” prompt, and assemble the system in the chat and the file pane. No terminal on this path. Claude drafts fast and computes with total confidence; you own every number that leaves the folder. A person ratifies every quality rule before it’s tested against, confirms every root cause before it’s reported as fact, and signs off on every figure before a stakeholder — or a second team — builds on it.
Capstone deliverable — Insight-in-a-Box (one brand, one analytics
function, end to end)
1. The foundation (Foundation / M1)
- dataset-profile.md — verified row count, real column meanings,
blank rate, duplicates, the "column that lies" catch
- metrics.md — 4-6 metrics, exact logic, named owner,
edge cases for each
- analysis-brief.md — one real request, decision it informs,
precise question, explicit scope
2. The query & join engine (Query / M2)
- one query that reconciles to a metrics.md figure, with a
plain-English walkthrough a stranger could audit
- one joined dataset: keys standardized and shown, unmatched rows
surfaced, counts and a financial total reconciled to both inputs
3. Quality & root cause (Quality / M3)
- data-quality-report.md — all six dimensions tested against
stated rules, each finding scored, an explicit go/no-go verdict
- root-cause.md — a real move sized against a baseline, the data
ruled out before the business was blamed, the cause confirmed
4. Experiments & charts (Decision / M4)
- experiment-readout.md — pre-registered hypothesis, the interval
not a bare percentage, guardrails checked, segments honest
- one verified, labeled chart — numbers reconciled before rendering
5. The reporting system (Report / M5)
- monthly-report.md — locked metrics, reconciled, charted, a plain-
English narrative with explicit caveats, a runbook to re-run it
- analytics-system/ — the query library, data contracts, CLAUDE.md,
monitor-spec.md, and a named owner with a real change process
Every piece must inherit #1. That inheritance is the whole assignment.The integration is not a formality. A reviewer will read the monthly report and check whether its MRR ties to the same definition metrics.md states. They will read the root-cause writeup and check whether it cites the same quality-audit finding that surfaced it. They will read the experiment readout and check whether its primary metric is one metrics.md already defines, not a fresh invention. Where those seams hold, the system works. Where they drift, the capstone flags a revise.
How it’s assessed — the master rubric
Five criteria, each scored meets / nearly / not yet. One “nearly” anywhere is a revise, not a pass — because an analytics function that reconciles in four places and drifts in the fifth still means a number reaches a board pack unchecked, and that is exactly when reconciliation matters most. The bar throughout is a single question: would Maya Haddad — or the most skeptical analytics lead you know — sign her name to this number, this query, and this verdict?
Master rubric — Insight-in-a-Box
Each criterion: meets / nearly / not yet. The bar: would a skeptical
analytics lead sign their name to this number, this query, this verdict?
1. Definitions are consistent everywhere
meets Every query, chart, and report cites metrics.md's exact
logic — no metric is redefined ad hoc in a later module. A
reader can't find two different "active" or "churn" logics
anywhere in the submission.
nearly Mostly consistent, but one artifact computes a metric
metrics.md already defines using a slightly different
filter, without flagging the deviation.
not yet Different pieces use different definitions for the same
named metric. The dictionary exists but isn't actually
the source of truth anything inherits from.
2. Every number reconciles
meets Every headline figure — in the profile, the query, the
quality scorecard, the experiment readout, and the
monthly report — ties to a second source, with the
comparison and the result both stated. Any gap was
investigated to a confirmed cause, not waved away.
nearly Most figures reconcile, but one number is stated without
a named second source, or a small gap is left
unexplained.
not yet Numbers are asserted without a reconciliation step
anywhere. A reviewer has no way to check whether a
figure is right short of re-deriving it from scratch.
3. Data safety
meets Customer-identifying data — names, emails, account-level
personal detail — stayed in the workspace throughout.
What was submitted is aggregates, patterns, and reconciled
totals; any ambiguous match (same entity, two ids) was
flagged for human review, never auto-merged.
nearly One artifact contains a real customer name or an
unflagged ambiguous merge that should have been routed
for review instead.
not yet Raw customer-identifying rows appear in a submitted
artifact, or an identity decision was made silently.
4. Arabic quality (bilingual teams)
meets Any Arabic narrative — the brief, a readout, the monthly
report — is authored in Arabic from the same reconciled
numbers, not translated from the English draft. Numbers
stay Western numerals; both languages cite one appendix.
nearly Competent but translation-flavored, or the Arabic and
English versions cite separately-built totals that could
drift.
not yet English run through translation. Reads foreign. Or the
Arabic pieces were omitted where they were called for.
5. Completeness
meets All five phases present and coherent. The quality audit's
finding is the root-cause investigation's starting point.
The query library inherits the dictionary. The monthly
report runs the whole pipeline on a schedule. A new
analyst could pick up this folder and reproduce any
number in it without asking you a question.
nearly A phase is present but disconnected — a quality audit
that doesn't feed the root-cause writeup it should have
triggered, or a monthly report that doesn't actually use
the M2 query library.
not yet Fewer than five phases, or phases present but not
connected to each other.Notice that four of the five criteria grade the seams between phases, not the quality of any single document. The modules already graded each artifact alone; the capstone grades whether the dictionary, the queries, the quality gate, the experiment readouts, and the recurring report hold together as one analytics function.
The five gates — what each module proved
The foundation (M1) proved that you can profile a file honestly and define a metric precisely enough that a stranger could turn it into a query without a follow-up question. The reviewer checks that dataset-profile.md reads real values instead of trusting column headers, that metrics.md states exact logic with a named owner per metric, and that analysis-brief.md names a decision and an explicit out-of-scope line, not just a restated question.
The query & join engine (M2) proved that you can produce a number a stranger could audit. The reviewer checks that every query’s WHERE clause cites metrics.md’s exact logic rather than a fresh guess, that the walkthrough is a real plain-English explanation and not a paraphrase, and that a join’s unmatched rows are surfaced with counts — not silently dropped — with both a row-count and a dollar total reconciled between the raw inputs and the joined output.
Quality & root cause (M3) proved that you can defend a number on demand. The reviewer checks that the quality scorecard tests all six dimensions against stated rules with severities, that the root-cause writeup rules out a data artifact — a missing date, a schema change, a drifted definition — before it’s allowed to name a business cause, and that the leading hypothesis names one confirming check that was actually run, not left as a plausible guess.
Experiments & charts (M4) proved that you can survive the second question. The reviewer checks that the experiment readout’s hypothesis and primary metric were pre-registered (or honestly flagged as exploratory), that the result is read as a confidence interval rather than a bare percentage, that guardrails were checked and any post-hoc segment is labeled exploratory rather than evidence, and that the chart’s underlying numbers were reconciled before it was rendered, not after.
The reporting system (M5) proved that you can turn one good analysis into infrastructure. The reviewer checks that the monthly report is genuinely repeatable — the same method, the same definitions, a runbook a teammate could follow — that every headline figure still reconciles on a recurring cadence, and that the self-serve system’s guardrails live in the CLAUDE.md and the query library themselves, not in the analyst’s head, with a named owner and a real change process.
Bring your own data
Use your own data. We recommend it for one blunt reason: the capstone stops being coursework and becomes two things at once — the actual analytics infrastructure your team runs on, and the credential that proves you built it to a professional standard. The metric dictionary, the query library, the quality contracts, the experiment-readout discipline, the recurring report — this is not a portfolio artifact, it is the system you open on Monday morning. That is day-one ROI: the assessment and the operational work are the same artifact.
The sample — Mizan’s analytics function — is here for when neutral ground is the point: a business analyst who wants to practice the integration once before rolling it out on live company data, or a team buyer who wants a worked reference to show the team what the output looks like before committing to the build. Either path clears the same rubric.
Sample brand: Mizan’s analytics function {#sample-brand-mizan}
Mizan is a GCC small-business bookkeeping SaaS with roughly 9,000 active customers, a Business Analyst — Maya Haddad — who supports Sales, Finance, Support, and leadership with numbers, and a Q1–Q2 2026 story that runs through every module in the track. Use it as your sample and the full arc is already threaded for you:
The foundation built around a real Monday-morning ask: Nadia Al-Sayed, Head of Customer Operations, wanted a churn dashboard. Maya profiled signups.csv (8,432 rows, a 12% blank source column, a plan column that turned out to record the requested trial tier rather than the current plan — a “column that lies” catch), defined Active Customers, Monthly Logo Churn, and MRR in metrics.md with Finance named as MRR’s owner, and turned the vague ask into a brief scoped to logo churn only, by plan tier and country, with revenue churn and a live dashboard explicitly out of scope for that round.
The query & join engine run on two real requests: active customers by plan, trailing six months, reconciling to 9,240 in March against metrics.md’s own figure independently — and a clean-and-join between customers.csv and orders.csv for Youssef Hamdan’s AR cross-check, which surfaced 64 orphan rows belonging to Najd Logistics under a legacy pre-migration id, and a near-duplicate Al Hosn Trading record flagged for human review instead of auto-merged. Every row and every dirham reconciled between input and output.
The quality and root-cause pass run the week Mizan’s March numbers were due in the board pack. The audit’s uniqueness check found 14 duplicate subscription_id rows, all annual-plan, all dated on or after March 1 — a BLOCKER. Its drift check independently caught an AED 9,400 MRR reconciliation gap. The root-cause walkthrough ruled out a tracking or schema artifact, localized the gap to annual-plan accounts charged on or after March 1, and confirmed via the billing-run job log that the March 1 annual-renewal batch fired twice — landing on the same 11 customers, the same trigger, the same date boundary that Mizan’s Support team found independently that same week via ticket triage. Two teams, two methods, one confirmed truth.
The experiments and charts module ran a 14-day vs. 30-day free-trial test (control 11.2% conversion, treatment 13.6%, a +2.4pp effect with a 95% CI of [+0.6pp, +4.2pp], guardrails clean) to a clean SHIP verdict for Starter-intent signups — and a verified weekly CSAT chart for the board, reconciled against CX Lead Layla Al-Nasser’s own tracking sheet, showing the honest six-week slide from 4.7 to 4.2 rather than a softened headline.
The reporting system assembled all four into Maya’s May 2026 recurring report — Active Customers up to 9,460, MRR at AED 508,000 (net-new +AED 11,000), and logo churn down to 1.2% as the Q2 save-flow from M1’s brief took hold, sharpest in the UAE Starter-tier segment the brief flagged — plus the standing self-serve system: a query library, data contracts, a CLAUDE.md, and a monitor watching, among other things, the exact billing double-charge pattern M3 root-caused, so it never needs a human to catch it manually again. That report’s MRR and logo-churn rows are what Youssef Hamdan pulls directly into Finance’s board pack — not recomputed, inherited.
Run all five phases on Mizan and the capstone story is complete: one brand, one chain of trust from a raw CSV to a number the board can act on. Then run it again on your own data, and the infrastructure is yours.
What the certificate honestly means today
Clear the rubric and you earn “Certified Data & Analysts with Claude” — a verifiable page at /cert/[id] you can share and add to LinkedIn. It is the credential of the team track: proof that you can run a complete analytics function on a coherent system, not just describe the moves.
Here is the honest state of it today:
- A person grades it, now. In a cohort or concierge run, a reviewer scores your analytics function against the master rubric above and tells you what is weak and what to revise. The rubric and the skeptical-analytics-lead bar are the product, and they are hand-gradeable today — that is the deliberate first step, not a placeholder.
- AI-assisted grading on the same rubric is next, not a claim we make yet. When it ships, it grades against this exact rubric — the bar does not move, the reviewer does.
- It is not an accreditation-body badge. Its weight is the work behind it: a complete, coherent analytics function a reviewer judged against a professional standard. That is a real thing to point at — and a more useful one than a quiz score.
No automated verification or paywall is wired behind this today that the line above does not describe. The certificate means what the work means.
What’s next
After the certificate, keep two things beside you as you run the system: the Data operating guide — the data-safety and verification layer that belongs underneath a team putting real company numbers through Claude — and the foundation toolkit from Module 1, the metrics.md and dataset-profile.md files that every later number inherits. When a definition changes or a new source comes online, update those two files and the whole system updates with them.
A manager dashboard — a team view of who is certified, which metrics are monitored, and reconciliation health across the function — is in the roadmap and is not a claim we make today. When it ships, it lands directly into the reporting system you built here.
The next capability track is your call. If your team also closes the books the numbers in this report feed, the Finance & Ops track covers the reconciliation and controls discipline from the finance side. If your team runs the queue the support tickets in this story came from, the Customer Support track covers the triage and quality loop on the other side of the same data.
Build the system, clear the rubric, and the certificate is yours — because the work is.