Performance advertising · AI systems · Attribution
I sit between marketing and engineering. I have bought the media and I write the code that automates it: running the spend, plumbing the data, putting a model where it earns its place, and checking the result against closed revenue.
The shape of the work
Campaigns built and staged through platform APIs rather than by hand in a UI, with performance pulled back into one tracker on a schedule. Includes running ads inside ChatGPT through the OpenAI Ads API, which is new enough that most of the work is reading the errors.
Built Meta's Conversions API on my own funnel: every server event carries an event ID matched to the browser pixel, so the pair deduplicates instead of counting twice. Offline conversions push from the CRM back to the platforms.
Most attribution problems are data problems. Source fields typed by hand, deals created by reps with no channel attached, stale numbers reaching a dashboard. I repair those on a schedule and show when the repair job stops.
Models draft the weekly summary, sort inbound messages by intent, and generate variants. Each one has a fallback so a failed call degrades the output rather than stopping the run.
Clicks and form fills are the easy numbers and they lie. I follow leads to whether they were real, and closed revenue back to the channel that started it.
On shipping LLM systems
The clearest example I have is a classifier that filters non-buyers out of paid search leads. It was not shipped because the output looked good. It was shipped because it ran against six months of past leads where the outcome was already known, disqualified about one in five, and blocked zero of the buyers who had gone on to real deals.
That is the pattern I apply to anything a model touches:
A prompt that works in a playground is a demo. The eval, the fallback, and the log are what make it a system.
Evidence
Meta Conversions API with pixel deduplication on my own funnel, and closed deals pushed from HubSpot back to Google Ads.
Read the case studyHand-typed source fields overwritten by tracking data every fifteen minutes, then closed revenue joined back to first touch.
Read the case studyA classifier backtested against six months of known outcomes before it was allowed near a live form.
Read the case studyCampaigns split by intent, one variable per test cell, and landing page work driven by reading every lead.
Read the case studyEight channels pulled automatically into one deck, with a model-written summary and a fallback when it fails.
Read the case studyA Rust dashboard over five systems that hands its context straight to an AI coding session.
Read the case studyClient build. Buying signals read from job ads and community posts, with every message human-approved.
Read the case studyMy own product: marketing data monitoring with an API that my other scripts run on.
Read the case studyHow I work
I am comfortable owning a system defined by an objective rather than a spec. What I insist on is that anything running unattended reports its own health, because the failure mode of automated marketing is not a crash. It is a job that quietly stopped in March and a number that has been wrong ever since.
Every scheduled job I build writes a log, and the ones that matter show a status light on a dashboard someone opens each week.
Confidentiality
The case studies describe methods and structures. Employer figures are rounded or given as change rather than absolute values. Account identifiers, campaign names, keywords, and anything about a named contact or deal are left out.
I will walk through any of it in detail in a conversation, at the level of method.