Deterministic-First Media Planning Automation

A media planning platform had campaign activation spread across manual planning, trafficking and reporting, with taxonomy errors surfacing late and no single system of record. I led the architecture and full-stack build, and ran the client discovery sessions myself.

The core decision: of fourteen pipeline stages, only six genuinely need language understanding (brief parsing, anomaly detection, reconciliation). The other eight are deterministic — and stayed that way. Letting a model orchestrate all fourteen “because it’s an AI project” would have been non-deterministic where the business needed determinism, harder to test, and impossible to cleanly audit. A rule like “apply taxonomy rule TX-014” is an if-statement, and an if-statement can be unit tested and explained to a client — it shouldn’t become a prompt.

Every LLM-backed stage has a deterministic fallback, so a model-provider outage degrades the system (less enrichment, more manual review) rather than stopping the business. The most transferable lesson from this project: client system access always arrives later than the build needs it — so I built every external connector with mock and real modes behind one interface from day one, meaning the pipeline was built, tested and demoable before real credentials existed.

Campaign activation went from weeks to under five days, with taxonomy exceptions surfacing at intake instead of at trafficking, and a full audit trail on every automated decision.