Experience

What I built here

• The CDP architecture, with the reason for each decision written down

  Architecture · AEP

  The obvious architecture does not work, and you have to measure it to find out. A journey writing to the profile leaves nothing in the data lake, so any state computed inside a journey needs nightly reconciliation to exist in SQL. Everything else follows: schemas, identity, and what persists where. With the reasons written down.

• Identity resolution with Data Prep and Query Service

  Data Prep · Query Service

  Identities resolved at ingestion with Data Prep and reconciled afterward through scheduled Query Service jobs. Duplicate profiles dropped by around 30% and unmapped emails went from 100% to under 50%.

• Consent from the event bus to profile and activation

  Kafka · AEP

  A consent pipeline from Kafka into AEP, with a separate load for the historical state that already existed. Both cadences matter: consent arriving as streaming and consent entering through a bulk load have to end up meaning the same thing.

• Weekly identity health runbook, with baselines

  Query Service · Identity

  Queries against the profile snapshot, run weekly and logged against a baseline: totals, identity combinations, and orphans by namespace. It is not a report, it is a threshold: if profiles carrying a single identifier climb too fast, the Data Prep mapping is wrong and you see it that week.

• Who not to delete: the hard part of a profile purge

  Data Lifecycle · Query Service

  Batch deletion of around 250,000 identities through Data Lifecycle work orders, until the licensed volume was back in range. The hard part is not deleting: it is choosing who not to delete. A candidate with an active policy or live personalization stays, and without that check the candidate list is worthless.

• Proving the purge happened, which is not the same as requesting it

  Data Lifecycle · Query Service

  Requesting deletion with full scope is not enough: the reconciliation dataset was not purged with it, and the nightly chain brought the profiles back every morning. Three independent queries surfaced that, and a second bug: the snapshot stores identity keys lowercase while events send them camelCase, so the checks returned false zeros.

• App personalization with recoverable operational state

  AJO · Operational state

  Personalization patterns in Adobe Journey Optimizer for the app, with the reconciliation needed to keep operational state auditable, recoverable, and transferable. Without that, a journey works until the first failure and nobody knows how to put it back.

• The campaign manager, inside AJO and without loops

  Journey Optimizer · Kafka

  The previous manager returned, in milliseconds, which incentive applied to each quote. Taking it apart showed a rules engine rather than a recommender, so Decisioning was not the answer for the first scope. The hard part was computing N incentives for N premiums with no loops in the expression editor.

What I built here

• OneTrust consent into AEP, by connector and by event forwarding

  OneTrust · Event Forwarding

  The official OneTrust connector ingests consent and preferences only, not cookies, and each run brings only what came after the previous one. It covers the preference center, not the consent that travels with a web event. I built both routes: the connector for identified data and event forwarding for the rest.

• AEP architecture across several business lines

  XDM · Identities

  XDM model, identities, Web SDK, and offer decisioning for separate business lines on one instance. The first line is never the problem. The problem is keeping the second one from forcing a rewrite of the first.

• The channel in AJO, with Pega untouched as the decision system

  AJO · Pega

  I coordinated the channel migration to Adobe Journey Optimizer without touching the decision system, integrating Pega with AEP and AJO, and consolidating more than 30 email campaigns. Changing channel and decision engine at once is the fastest way to lose track of what broke.

• Journey Optimizer Experience Decisioning in banking

  Experience Decisioning · Journey Optimizer

  Work on Experience Decisioning, the current decisioning capability in Adobe Journey Optimizer, rather than the Decision Management inherited from the earlier offer decisioning. The difference changes where an offer lives, how it is requested, and what the channel can decide in the moment.

• Profile or context attributes: where each value comes from

  Journey Optimizer · Personalization

  Personalizing a send has two sources, and picking the wrong one is expensive. What the CDP already knows about the customer comes from the profile. What is only known at trigger time, and therefore cannot be anticipated, travels as a context attribute in the call itself. I defined which of the two feeds each value.

• Retargeting with viewed offers kept in the browser

  Decisioning · GTM

  Retargeting offers without standing up a backend. The last ones the user viewed are kept in the browser, in a short queue fed by the card selection event. They come back with the next decisioning request, so the response does not repeat itself.

What I built here

• Shared data layer for web, logged-in area and app

  Data layer · Web and app

  Three surfaces with three separate implementations, and therefore three versions of the truth. I rebuilt the data layer on all three against one event contract, coordinating a team of 5, with functional and technical documentation kept in Confluence.

• Walmeric and Glassbox wired into the measurement layer

  Walmeric · Glassbox

  Phone lead capture in the logged-in area ran through Walmeric. I integrated the form and its submission, and carried the Walmeric visitor identifier into the data layer, which is what ties a call back to the session behind it. On Glassbox, I improved its integration with Adobe Analytics.

• The risk map of the tag manager

  Tealium iQ · Audit

  You audit what you inherit, not what you built. A tag manager can rewrite the DOM and download scripts, so I ranked by risk what each piece could do across both sites. With a cookie inventory, Adobe Analytics cross-domain tracking, and the library served through a reverse proxy on its own domain.

• Replacing a bank tag manager without switching measurement off

  Tealium iQ · Data layer

  I inherited it from another supplier and it could not be trusted, so nearly all of it was rebuilt: the data layer, the load rules, the extensions and the tags. All of it with the site in production and no downtime window, which forces you to move in pieces and prove the new one says what the old one said.

• Two Tealium profiles inside one brand

  Tealium iQ · Adobe Analytics

  One brand, two different worlds. The commercial site and the transactional one share no tags, no consent rules and no risk profile. I kept a profile for each, with Adobe Analytics and Target in both, and most advertising tags only in the commercial one.

• Custom campaign channel taxonomy in Adobe Analytics

  Adobe Analytics · Attribution

  Campaign parameters in the URL are parsed with a per-channel grammar and mapped into Adobe Analytics variables, each with a declared meaning. It replaced an earlier position-based parser that broke as soon as an agency changed its tagging format.

• Server-side tracking and paid media attribution

  EventStream · Server-side GTM

  Paid media pixels deployed client-side and server-side, with routing on Tealium EventStream and server-side GTM. Identifiers left the browser hashed, validated before hashing and only where consent allowed. The platform processed tens of millions of events a month.

• Cross-domain consent with Didomi, from scratch

  Didomi · CMP

  Cross-domain consent: a decision taken on one domain had to hold on the others and survive the jump. Pre-production and production, web and app webview, categorized purposes, custom banners. And that decision reaching everything: the TCF API for vendors, Consent Mode v2 for Google tags.

• Measurement held until the user decides in the banner

  Consent · Tealium

  The first page view happens before anyone has accepted anything. While the banner is open the page sits behind a cover, and that view is stored rather than lost or sent regardless: it is replayed afterward, following the accepted categories. Without consent nothing leaves, and with it the entry to the session survives.

• Promotional component measured inside the logged-in area

  Adobe Analytics · Logged-in area

  A promotional component inside the authenticated area, with its placement and its click measured as their own variables rather than as one more navigation. Each position carries its own slot, which is what allows placements to be compared and then kept or pulled on evidence.

• What session replay is allowed to record

  Glassbox · Privacy

  A session replay tool sees the user screen, and inside a bank logged-in area that screen is their money. Rather than trusting default masking, what it may capture is declared in a list: outside it, nothing is recorded. Installing the tool is easy. Deciding what it has a right to see is the work.

• Adobe Target as a CDP destination, with audiences governed

  Adobe Target · Real-Time CDP

  Target not as an island but as a CDP activation destination with edge personalization. Publishing audiences from analytics has a cap and takes hours to become actionable, so the rule was written the other way around: audiences are created in the CDP and never in the personalization tool.

• One collection layer, on the Adobe Web SDK

  Adobe Web SDK · alloy.js

  Moving from legacy per-surface libraries to a single collection layer on the Web SDK. This is the piece that makes everything after it possible: without one collection layer, server-side, the unified profile, and the move to CJA all inherit the divergences they were meant to fix.

• Adobe Analytics migrated to Customer Journey Analytics

  Adobe Analytics · CJA

  With the SDR maintained by hand for years, its eVars, props, and events, and the data quality control that held the reports up. The migration was not copying the old model across, but deciding which part of it still supported decisions.

What I built here

• Measurement template containers, reused across accounts

  Adobe Launch · GTM

  Reusable GTM template containers, exported and imported into each client account: one measurement base, one for ecommerce, one per consent platform, and one for app. Standardizing retail, fashion, travel, and automotive forces you to separate the data contract from the tool that implements it.

• Early migration from Universal Analytics to GA4

  GA4 · Universal Analytics

  For a fashion account, with the platform just released and none of the guidance that exists today. Both properties ran side by side in the same container through the transition, which is what lets you compare numbers and explain the differences before switching the old one off.

• Analytics SDKs on iOS and Android, and hybrid integrations

  iOS · Android

  Native SDK implementation and resolution of the hybrid integrations between the native layer and the embedded web view. That is where sessions and identities get lost, and nobody looks until the numbers stop matching.

• Daily SFTP ingestion into Drive, with three update modes

  Google Cloud · Python

  A Cloud Function starts a virtual machine at the scheduled time; the machine reaches the SFTP, loads the CSV files into Drive and Sheets, and shuts itself down when finished. The machine exists because of one specific constraint: the server only accepted a fixed IP, and Cloud Functions had none.

• Notebooks that configure analytics in bulk and close the ticket

  Python · Management API

  Python notebooks against the Google Analytics management API to create and update goals, filters, custom dimensions, and permissions in bulk across accounts, properties, and views. Every change was diffed against the live configuration first, and the Jira ticket was transitioned on completion.

• Server-side tagging with custom clients, when sGTM had just shipped

  Server-side GTM · Custom clients

  A server-side container with custom clients, not just the stock ones. One for the legacy collection, one for the new one, one for an ad platform, plus outbound requests to an in-house API. The container was exported in March 2021, with the tool barely out.

• Dashboards per market, with access governed

  Reporting · Governance

  A consumer goods brand with subsidiaries across several European markets, and one dashboard for each. Building them was not the hard part. Deciding who sees what was, so every market carried a declared recipient and access level. A report shared too widely leaks and one shared too narrowly goes unused.