Building autonomous AI infrastructure for martech and paid media teams. Paid-Media-Agent is a closed-loop system where agents act as forensic data analysts, causal modelers, and programmatic media buyers, backed by rigorous analysis and strict enterprise data governance.

An enterprise-grade AI operating system for paid media. Three autonomous agents handle forensic data auditing, causal attribution modeling, and programmatic budget execution. An MCP server connects Claude directly to live BigQuery data and agent outputs. A set of Claude Code skills handles strategy, planning, and ad-hoc analysis — all inside your own GCP environment, zero data outflow.

graph TD
    human["👤 You + Claude"]

    mcp["paid-media-mcp<br/>MCP server · TypeScript"]
    skills["skills<br/>Claude Code slash commands<br/>/paid-media/*"]

    watchdog["Watchdog<br/>forensic audit · signal capture · anomaly traps"]
    analyst["Analyst<br/>BSTS · Meridian MMM · Shapley · Markov"]
    operator["Operator<br/>budget mutations · audience suppression<br/>±10% guardrail · all-or-nothing gate"]

    schema["BigQuery Schema<br/>17 SQL layers · identity namespace registry<br/>inside paid-media-agent/schema/"]
    platforms["Platform APIs<br/>Google Ads · TikTok · Meta · LinkedIn · Reddit · GMP"]

    human --> mcp & skills
    mcp --> watchdog & analyst & operator
    watchdog & analyst & operator --> schema
    operator --> platforms
    platforms -.->|daily exports| schema

AI-powered privacy compliance monitoring for web properties. A headless browser worker captures every third-party tracking script firing on a website, an LLM parses the site's privacy policy to extract disclosed vendors, and a compliance engine compares the two — flagging undisclosed trackers, PII leakage, and policy mismatches with severity-graded violations and AI-generated remediation steps.

graph TD
    scheduler["Cloud Scheduler<br/>hourly cron"]
    cron["/api/cron/trigger-scans<br/>checks scan frequencies"]
    queue["Cloud Tasks Queue"]
    worker["Playwright Worker<br/>headless Chromium"]

    capture["Network Capture<br/>intercept all outbound requests<br/>17 vendor fingerprints"]
    policy["Policy Extractor<br/>find privacy link · extract text"]

    chain1["LLM Chain 1<br/>extract disclosed vendors"]
    chain2["LLM Chain 2<br/>detect PII in payloads"]
    chain3["LLM Chain 3<br/>compare observed vs disclosed"]

    dashboard["Dashboard<br/>health score · violations · reports"]
    db["PostgreSQL<br/>Prisma ORM"]

    scheduler --> cron --> queue --> worker
    worker --> capture & policy
    capture & policy --> chain1 & chain2
    chain1 & chain2 --> chain3
    chain3 --> db --> dashboard

Data Integrity — Forensic State Machine Catches Salesforce CRM attribution overwrites before they reach models. Fingerprints systemmodstamp + lead_source_updated_at drift to surface bulk imports that silently rewrite "Google Ads click" to "Trade Show Badge Scan." Also detects organic traffic surge patterns that mask true paid media incrementality.

Causal Measurement — JAX-backed BSTS Bayesian Structural Time Series counterfactual modeling. Isolates true paid media lift from organic trend with credible-interval confidence bands — not correlation, actual causality.

Portfolio Allocation — Google Meridian MMM Full Bayesian Marketing Mix Model. Posterior ROI distributions + confidence tiers generate task27.v1 budget recommendation packages consumed directly by the Operator agent.

Programmatic Write — All-or-nothing mutation gates Sequential budget mutations across Google Ads, TikTok, Meta, LinkedIn, and Reddit. Pre-flight guardrail sweep validates schema version, human approval flag, ±10% shift cap, and platform floors — zero mutations if any check fails.

The Attribution Heist Pipeline review shows 40% of last quarter's closed-won attributed to "Trade Shows" — but no events ran. The Watchdog catches it: 150 CRM records where systemmodstamp = lead_source_updated_at, all timestamped five days after the original paid click sessions. A Salesforce bulk import overwrote the gclid attribution.

The "Organic Spike" That Isn't Organic sessions spike 5× for five days. Marketing declares a win. The BSTS counterfactual says otherwise — paid spend was flat, there's no organic driver in the causal graph, and the credible interval collapses around zero incremental lift.

Quarterly Budget Reallocation Meridian MMM shows LinkedIn at 3.1× adjusted ROI vs. TikTok at 1.85×. The Operator generates a task27.v1 package, runs the pre-flight sweep, and executes across all five platforms in one sequential pass — capped at ±10%, full audit trail in BigQuery.