The marketer who builds.

I am Robi Powers, a GTM engineer. I build AI agents, outbound automation, and micro‑apps into revenue operations. I have shipped dozens of systems. This page documents six, the ones that matter most for revenue work, with real costs and real results.

One of them is on shift right now: a WhatsApp coordinator that manages a video production client’s feedback around the clock, for about $17 a month.

I work remote on US time zones. I am a US citizen, based between Los Angeles and San Salvador, and I deliver in English and Spanish. The fastest way to reach me is robipowers@gmail.com.

System 01 · The daily prospect engine

Auction Mentor, my own subscription membership business, needed pipeline without spam. I built the engine to run on a weekday schedule and hand a human a queue, never a send button.

1. 01

   Source.

   Apify scrapers pull qualified prospects from intent signals: hashtag activity and self‑identification in bios.

2. 02

   Score.

   Claude grades each profile against a written ICP and discards the rest.

3. 03

   Dedupe.

   A queue state machine checks every prospect against all prior runs.

4. 04

   Draft.

   The engine writes a personalized, value‑first DM for each qualified prospect.

5. 05

   Approve.

   A human reads, edits, and sends every message. The approval gate is a design decision, not a limitation.

It was built with a hard $0.50 per‑run cost ceiling. Its live production run queued 20 qualified prospects for $0.38.

Apify · Claude · Node.js · scheduled automation

Rebuilt in Clay, July 2026. I build these systems in code. This one I rebuilt in Clay, the platform GTM teams actually run on, in a single day: the same qualification logic, rule for rule, aimed at a new intent signal, companies hiring GTM engineers. Waterfall enrichment, Claygent research with cited sources, deterministic scoring. The rigor came with it. When an enrichment source returned the wrong company, the human review step caught it, and where data was thin the system returns an honest blank instead of a guess.

Clay · Claygent · waterfall enrichment · AI scoring columns

System 02 · The coordinator that never sleeps

Garage Films, a video post‑production company, ran client feedback through one overloaded human. The replacement works the thread around the clock. It triages replies, converts feedback into structured tasks for editors, and verifies every delivered render with audio and vision QA before anything goes back to the client.

System 03 · The 16‑minute package generator

The Client Operating System is a productized offer: a done‑with‑you AI workspace for small businesses, sold at $300, $500, and from $900. The delivery problem was consistency. The answer is a generator built on a 20‑agent Claude workflow. It takes 4 intake inputs and produces the complete deliverable, 16 to 17 files in Spanish, in about 16 minutes. One rule is wired in: no fabricated numbers, ever. It has delivered for 9 client businesses in El Salvador.

Node.js · Claude multi‑agent workflow · delivered in Spanish

System 04 · Attribution without guesswork

Paid acquisition for that offer needed clean numbers end to end. When a customer pays, a Stripe webhook fires: the signature is verified, customer data is hashed, the event is deduplicated, and a purchase lands in Meta’s Conversions API. The ad account learns from real revenue, not proxy clicks. Around the bridge sits the rest of the stack: bilingual React landing pages on Vercel, Meta Pixel events verified live, and an 18‑asset static ad batch rendered through a reusable HTML‑to‑PNG system.

Stripe · Meta Conversions API · React · Vite · Vercel

System 05 · A product with the agents built in

Auction Mentor (app.auctionmentor.io) is my own product: a live subscription membership business at $150 a month, built on Stripe, Supabase, and Discord. Members work with a proprietary indicator suite and an AI copilot that reads their live third‑party dashboards. The copilot works through a 68‑tool MCP server written in Node.js on the Chrome DevTools Protocol. Every weekday morning a scheduled pipeline synthesizes an AI analysis and broadcasts it to the production database every member’s copilot reads. A read‑only watchdog snapshots the operation’s health daily. It caught a real 5‑day silent outage on its first run.

Stripe · Supabase · Discord · Pine Script v6 · Model Context Protocol

System 06 · The enterprise engagement

Grupo Q, an automotive group with 67 locations across 6 countries, needed to know what its customers actually said. The engagement ran in two phases. Phase one listened. Apify scrapers pulled 11,257 public reviews from across the group’s locations, and a 23‑agent workflow clustered them into a customer‑voice census with research‑verified positioning. Phase two armed the team. Their marketers now hold a brand‑locked Claude plugin, a router and five skills, covering campaign creative, per‑country localization and currency handling, social calendars, and financing copy with legal guardrails wired in. The census earned a direct job offer from the company’s president.

Apify · Claude multi‑agent workflow · brand‑locked Claude plugin · delivered in Spanish

The new channel

One more capability belongs in this record. Buyers increasingly ask ChatGPT and Perplexity instead of Google, and being the source those engines cite is a discipline of its own: AEO and GEO, answer engine and generative engine optimization. I run it as a service today. My audit tooling scores citability, AI crawler access, llms.txt coverage, schema markup, and brand‑mention presence across the AI platforms. Clients get it as a scored report, alongside traditional SEO.

GEO audits · citability scoring · llms.txt · schema · client‑ready reports

The record