Kayden Pellegrini

The problem

The business could not say what it actually made on a product. Margin was being worked out from list price and valuation figures, and neither of those is what anything cost.

Stock bought abroad arrives on a purchase order in a foreign currency, shares the freight, customs and bank charges of the shipment it came in on, and is sold one numbered unit at a time. A list price captures none of that.

What I built

Landed cost rebuilt at the grain of the individual serialised unit. Each unit takes its cost from the supplier price at the exchange rate actually paid on that transaction, plus its share of the freight, customs and bank charges on its shipment.

Every inventory record traces back to the purchase order line that paid for it, so any margin figure can be followed to its source. The logic runs through a mix of Dataverse and dbt models, with tests written to fail the build when a number stops reconciling. A figure that cannot be defended stops there instead of reaching a report.

What changed

Margin by product is measured against real cost rather than list price or valuation figures, and any figure can be followed back to the transactions that produced it.

The problem

Stock was counted by hand across three provinces. Counts were slow and hard to trust, and when one finished there was no clean way to separate what had been confirmed from what was missing and what had turned up where it was not expected.

What I built

Handheld RFID scanners feeding Power Apps and Dataverse, with Power Automate producing HTML and printable reports at the end of each count.

The off-the-shelf capture path was too slow, so I wrote a native iOS app in Swift that decodes SGTIN-96 EPC tags and GS1 barcodes over Bluetooth Low Energy and passes serials straight into the stocktake app. Reconciliation is part of the count itself rather than a job for afterwards.

What changed

A 30 item scan went from minutes to under 15 seconds. Every count reconciles confirmed, missing and unexpected stock in one pass, across all three provinces.

The problem

AI-generated changes to a real system cannot be taken at face value. A model will report a fix it did not make, rename something and call it a security improvement, or claim a control exists because the word appears in a comment. "It says it fixed it" is not evidence.

What I built

A multi-agent verification framework. Independent passes review and then confirm each change, and the confirming pass never sees the reasoning of the reviewing pass. An adjudication pass then checks every claimed change against the actual result and classifies it as verified, overclaimed, hallucinated, cosmetic or a regression.

The LLM tooling itself reaches Dataverse, Power Apps and Microsoft 365 through Model Context Protocol, so analysis runs against the live systems instead of a pasted extract.

Around it sit the standards I set for AI-assisted analysis on live systems: one source of truth per data domain, confidence labelling, read-only scoping, human confirmation before anything destructive runs, no merging of records that have not been verified, and no margin calculated from a placeholder cost.

What changed

A repeatable way to trust or reject AI output on production systems, decided by what actually changed rather than by what the model says changed.