I build and run data systems.

I implemented a customer data platform as the central event pipeline and wired it into the Google Ads Conversions API with enhanced conversions: server-side purchase events carrying order id, revenue and subscription tier, with hashed identifiers for match rate. Later a BigQuery layer for checkout events and attribution. Roughly 75,000 tracked users a month, in production since 2023 and still running.

I fed item-level cost of goods from the product feed into the platform so bidding optimises on gross profit directly. That meant auditing how the margin was calculated before trusting it, building the supplemental feed for both markets, and fixing a rate limit on the client API that was silently rewriting conversion values. First fourteen days after launch: 2,737 EUR spend, 11,593 EUR gross profit, profit ROAS 4.23.

Found by validating response bodies rather than status codes. A 200 with the wrong body is invisible to every monitor that watches status alone. The same class of failure turned up at a second client three weeks later, which is why I now check payloads by default.

I pulled orders from the shop database and from analytics and compared a full month before against a full month after. The share had not risen, it had fallen. Reconciling against the source system rather than against a second measurement is the only way that result surfaces, and reporting it was worth more to the client than defending the deployment.

The argument was being had in screenshots, so I brought warehouse figures into their own comparison sheet, reconciled the discrepancies and explained in plain terms how that data feeds back into analytics to improve accuracy. The deliverable was not a tag, it was a client who could see what they were paying for.

A control script reported the state of each consent signal instead of leaving it to inspection of the Tag Manager screen, which showed which signals never reached granted. The fix went across multiple domains under one container. Consent is where most measurement breaks now, and it breaks silently.

I automated the whole chain: leads from the web into the system, the responsible person assigned automatically, a duplicate check that runs before a record is created rather than after, email history attached to the client, calendar integration and reporting into the BI tool. Thirty-one shipped tasks over the project.

Ads paused ahead of the switch, analytics rebuilt on the new platform, and the data reconciled so the series before and after line up rather than showing a cliff. The gap is normally found months later when somebody questions a report.