An archive of the dashboards, models and reports I've actually shipped.

Four ad platforms each claiming the same conversions. Weekly reporting took six hours of copy-paste and still nobody trusted the total. Spend decisions were being made on whichever dashboard was open at the time.

A single warehouse model joining platform spend to order-level revenue, with a documented last-touch and a position-based view side by side. One dashboard, one definition of a conversion, refreshed nightly.

The monthly board pack took three days of manual assembly, and the runway number changed depending on who built it. Finance was spending its month closing the last one instead of reading the current one.

A modelled P&L off the accounting system, a burn and runway calculation agreed in writing with the CFO, and a Power BI page that exports straight into the board deck. Reconciled to the ledger to the riyal.

Churn was reported as a single monthly percentage, which told the team nothing about who was leaving or when the decision to leave was actually made. Every save attempt happened after the cancellation.

Monthly cohort retention tables in Tableau, plus a Python scoring job that flags four leading behavioural signals in the 30 days before a typical cancellation. Flagged accounts land in the CS team's queue automatically.

On-time delivery was reported at 96% while customer complaints said otherwise. The SLA clock started at dispatch, not at order, so every delay before the van left the warehouse was invisible.

Re-defined the SLA from order timestamp, rebuilt the funnel stage by stage, and put cycle time per stage on one Tableau view. The number dropped to 87% on day one, and for the first time it matched what customers were experiencing.

A retail group was two weeks from signing a six-figure warehouse migration. They asked me to audit the current stack first, mostly as a formality.

Three of the four reported problems were tracking bugs, not capacity limits. Two were duplicate event fires inflating traffic by 31%. The migration was postponed; the fixes took eleven days and cost a fraction of the quote.

Branch performance was measured on deposits and digital on logins, so the two channels were never comparable. Nobody could answer what a customer acquired in a branch was actually worth against one acquired in the app.

A customer-level profitability model covering revenue, cost-to-serve and product holding per customer, attributed back to acquisition channel. One view that the branch network and the digital team both agreed to be measured on.

Nine properties, nine spreadsheets and three different definitions of RevPAR. Revenue managers were pricing against numbers that disagreed with what head office was reading in the same week.

A single property-night grain model feeding a group Power BI view: occupancy, ADR, RevPAR and booking pace against the same period last year, per property and consolidated.

Route decisions were being made on load factor alone, so full aircraft on unprofitable routes looked like wins. Ancillary revenue and per-route cost never appeared in the same table as the seats sold.

A route contribution model covering fare, ancillary, fuel, crew and slot cost by cabin, with a Tableau view that ranks routes by contribution rather than by how full they look.

Ticketing, merchandise, hospitality and food service all reported separately. The club knew precisely what a match earned and had no idea what a fan was worth across a season.

A fan-level view joining ticketing, retail and hospitality spend, plus a renewal model that flags at-risk season-ticket holders before the renewal window opens rather than after it closes.

Every office reported its own average processing time, every office met target, and citizen complaints kept climbing. The averages were hiding the tail entirely.

A percentile-based SLA model built on median, P90 and P95 instead of the mean, with a plain-language summary view. The P90 told a very different story, and it was the one that matched the complaints.

Category managers were negotiating with sellers using last month’s spreadsheet. Slow-moving SKUs were only spotted once they had already tied up working capital for a full quarter.

A weekly-refreshed sell-through and seller scorecard covering days of cover, return rate, fulfilment reliability and margin per seller, with an automatic flag when a SKU crosses its cover threshold.

Motor claims were assessed one file at a time, by different adjusters across three regions. Duplicate invoices, inflated repair estimates and repeat claimants were only ever caught by chance, and never counted.

A claims model joining policy, claimant, workshop and invoice history into one record, plus a nightly scoring job that ranks incoming files by leakage risk. The top decile routes to a senior adjuster before payment is released; everything else clears automatically.