PMAX — Total Marketing Partner

Business teams had plenty of automation ideas but no way to turn them into reliable agents without an engineer in the loop.

I conceived, architected and built OSAF end to end: a Claude Code/Codex plugin with 14 commands, 22 playbooks and a 100-point evidence-based QC rubric, plus a deployment portal on Next.js, Prisma and Redis.

I drove adoption by consulting business users on their use cases and supporting their deployments. 10 agents have shipped to production with OSAF support, including some built before the portal existed, and since its launch the portal has onboarded 18 users and 14 agent projects from 10 teams.

Live ad setups on Meta, TikTok and Google Ads drift away from the media plans in Google Sheets, and catching that by hand does not scale.

As top contributor, I co-built an end-to-end QC service (NestJS, PostgreSQL/Prisma, AWS ECS) and its React UI that validates live setups against media plans on cron schedules. Its 69 rules and 184 checks have run 62K+ times, sending mismatch alerts to account teams via Gmail and Google Chat and flagging broken source sheets before they corrupt results.

When database transaction and query timeouts started failing scheduled runs, I traced and fixed them, cutting failed runs by 79% within a month and isolating the rest to faulty input sheets. I also added on-demand runs, rate-limit retries and S3 archiving.

On Airflow (Python), built a missing-permission monitor that alerts account owners and auto-deactivates accounts after three reminders. It resolves 97% of flagged accounts (about 60% restored, the rest switched off) within a median of 3 days.

Reworked the ad-account-to-project mapping flow across Airflow DAGs, a DAG-trigger service, the client API and the React UI: auto-mapping configs, separate manual and hourly run status, and project-ID revalidation with alerts for unmapped projects. The pipeline keeps ~98% of live projects refreshed within 2 hours.

Also shipped a daily 24-hour optimization report spanning QC, alert, performance and trend rules, plus alerting and validation for CRM/KPI uploads.

Extended the performance-control and alert rule engines, which have run 500K+ times, with action status and suggestion reporting plus failure alerts, and integrated the TikTok Smart+ campaign API.

Migrated an ad-compliance auditing product's LLM pipeline from OpenAI to AWS Bedrock, and built its web-crawler compliance scan and asset monitoring.