GTM Engineer · AI Marketing Systems
The marketer who builds.
I build production AI and revenue systems with measurable outcomes and verifier-driven quality.
I am Robi Powers, a GTM engineer based in Los Angeles. I design the loops behind revenue work: AI agents, routing, attribution, customer workflows, and the checks that keep them reliable.
- 70 agents
- across 4 reviewed production rounds
- 6 pages
- shipped to the reference bar in under 3 hours
- 82 tests
- behind a 24/7 customer workflow
- 192,000
- Instagram audience built from zero
Selected production systems
Seven systems, each documented with the build, the control layer, and the result. The throughline is simple: ship the workflow, encode the judgment, and measure what changed.
System 01 · The production loop with its own critics
I designed and ran a multi-agent production pipeline in Claude Code: 70 agents across 4 rounds, with parallel builder agents working under disjoint file ownership. Blind reviewer agents captured rendered output and failed anything below the reference bar set by Linear, Stripe, and Vercel. Specialist audit agents checked accessibility, tablet breakpoints, SEO, and performance.
| Build | parallel agents with explicit ownership boundaries |
|---|---|
| Review | blind screenshot comparison against named reference sites |
| Verification | accessibility, breakpoint, SEO, and performance audits |
| Result | a complete 6-page marketing site cleared review in under 3 hours |
Claude Code · agent orchestration · loop engineering · verifier design
System 02 · The coordinator that never sleeps
One of Central America's largest production companies ran client feedback through one overloaded human. I shipped a 24/7 WhatsApp AI coordinator that triages replies, converts feedback into structured tasks for editors, and verifies delivered renders with audio and vision QA before anything returns to the client.
| Status | live in production, operating around the clock |
|---|---|
| Cost to serve | roughly $17 per month |
| Documented margin | roughly 94% |
| Quality layer | 82 tests plus audio and vision verification |
WhatsApp · Claude · DigitalOcean · deterministic QA
System 03 · Enterprise customer intelligence
One of Central America's largest automotive groups, operating 67 locations across 6 countries, needed a reliable view of what customers actually said. I analyzed 11,257 public reviews with a 23-agent workflow and a 112-check deterministic audit, then translated the evidence into executive priorities and research-verified positioning. The engagement earned a direct role offer from the group president.
| Coverage | 11,257 reviews across 67 locations and 6 countries |
|---|---|
| System | 23 research and synthesis agents |
| Audit | 112 deterministic checks |
| Outcome | executive-ready findings and a direct role offer |
Apify · Claude multi-agent workflow · customer research · delivered in Spanish
System 04 · Attribution without guesswork
I built a paid-acquisition stack that connects bilingual React and Vite landing pages, Stripe Payment Links, Meta Pixel, and a server-side Stripe-to-Meta Conversions API bridge. Payment signatures are verified, customer data is hashed, events are deduplicated, and the ad account receives confirmed revenue rather than proxy clicks. The launch included 18 campaign assets rendered through a reusable production system.
Stripe · Meta Conversions API · React · Vite · Vercel
System 05 · The 16-minute package generator
I commercialized a Node.js and 20-agent Claude workflow that turns 4 client inputs into a complete 16 to 17 file AI workspace in about 16 minutes. The generator enforces a zero-fabrication rule and has delivered across 9 paying client businesses, replacing a variable manual process with a documented and repeatable operating system.
Node.js · Claude multi-agent workflow · structured outputs · delivered in Spanish
System 06 · The scoring and approval engine
I built a signal workflow with Apify ingestion, enrichment, written ICP criteria, deterministic scoring, deduplication, cost controls, and a human approval queue. It produced 20 qualified prospects for $0.38 under a hard $0.50 ceiling. I later rebuilt the same qualification logic in Clay as a portfolio demo, preserving the rules and the review gates rather than treating the platform as a shortcut.
| Clay funnel | 71 companies sourced, 38 qualified against the rubric |
|---|---|
| Enrichment | two 4-provider waterfalls for job openings and funding |
| Research | Claygent research with cited sources |
| Cost control | 443 credits, below the self-imposed 800-credit ceiling |
| External action | none; every record remains behind human review |
Clay · Claygent · Apify · waterfall enrichment · deterministic scoring
System 07 · A product with the agents built in
I founded and built a live subscription product on Stripe, Supabase, and Discord. Its AI copilot reads live third-party dashboards through a 68-tool MCP server written in Node.js on the Chrome DevTools Protocol. A scheduled pipeline writes shared analysis into the production database, and a read-only watchdog records structured health verdicts. The watchdog surfaced a real 5-day silent pipeline outage.
Node.js · Model Context Protocol · Chrome DevTools Protocol · Supabase · Stripe