Open to opportunities

Rafael Rabelo

I buildAI agentsCRM automationsemantic searchAI web appsfunnel dashboardscomputer visionin production

AI Ops / Automation Engineer — AI Agents · Workflow Automation · GTM Tooling

Let's work together Recruiter? Download CV rafaelrabel0

7+: B2B clients served
20+: Accounts onboarded
40+: Skills & tools
C2: English (EF SET)

About

AI Automation Engineer and founder of Rabelo Co., building production AI agents, automation workflows, and data pipelines that turn business context into measurable revenue tools. Hands-on across the full stack: n8n agent orchestration, LLM integration (Claude, GPT), RAG with vector databases, CRM automation, and AI-powered web apps shipped end to end. I sit at the intersection of sales, product, and AI engineering. English C2 (EF SET Certified).

Location

Remote · Brazil

Languages

Portuguese (Native) · English (C2 Proficient (EF SET))

Education

B.Sc. Software Engineering — In progress (2024 – Dez 2027)

Certifications

EF SET English Certificate — C2 Proficient (77/100), May 2024 · Technical training in Automation & AI — Escola de Automação, NoCode Academy, Kommetrik

Experience

The path: from standalone agents to production platforms.

Apr 2025 — start

Rabelo Co. — from sales into engineering

First production AI agents on n8n, wiring Claude/GPT into WhatsApp through the Evolution API: consultative qualification, follow-up and human handoff with live state.
Time spent in sales became the technical edge: qualification criteria and tone of voice come from the real operation, not a generic prompt.

2025 – 2026

GTM suite and funnel observability

Stage-aware activity cadence engine in Pipedrive (128-node workflow) creating touchpoints per stage and date — and rolling activities back when a deal changes stage.
Automated post-call intelligence: Whisper transcription of the SDR call → GPT summary → note logged on the deal.
Event-fed funnel dashboard through a per-client authenticated edge function — each new agent plugs in with three emitter workflows.

H1 2026

Hybrid RAG and in-house products

RAG + API architecture (Supabase pgvector, 1536-dim embeddings) serving static knowledge from the vector store and reserving live calls for scheduling — migrated 20 professionals, 224 services and 500 clients.
In-house products shipped: a full personal-finance web app (Next.js, Supabase with RLS, REST API with keys) and a management PWA for 3D printing.
Semantic search engine over 1,266 indexed 3D-model libraries, fusing fuzzy and vector search with GPT-4o Vision visual verification.

H2 2026

SaaS platforms and computer vision

Content approval and publishing platform in production (Next.js 16 + Supabase + Meta Graph API): the client approves through a public link and the app schedules and publishes to the network.
Multi-tenant CRM with an AI agent inside the same platform — public chat, lead capture and CFM/LGPD guardrails, with per-conversation cost measured against the API before shipping.
Face-recognition search in production: 3,995 photos indexed, 15,681 faces, ONNX Runtime on GPU and pgvector with an HNSW index — send a selfie, get your photos back in under a second.
AI virtual assistant with a live demo (n8n + Whisper + GPT), handling text, audio and images to qualify and schedule.

Selected work

Point at a project for the screenshot; click to open the write-up.

AI agents

Agente SDR WhatsApp MultimodalAtivos Digitais

● Live

AI-powered WhatsApp service, qualification and follow-up — 24/7

Problem: Leads arriving on WhatsApp at any hour with no qualification, no CRM record and no consistent follow-up — pipeline silently evaporating.
Solution: Multimodal n8n agent (text, audio via Whisper, image via Vision) with a Redis message buffer, Postgres chat memory and Supabase lead registry. Qualifies consultatively, moves Kommo stages and tags, uses dedicated tools (qualification, name update, human handoff) and emits funnel events to the dashboard. Cascading follow-ups (10 min → 14 days) auto-close silent conversations and resume them when the lead replies.
Result: Live across multiple client operations; 24/7 service with AI→human handoff and an end-to-end tracked funnel.

24/7Coverage

9Orchestrated workflows

107Nodes in the main agent

Stack

n8n (9 workflows)OpenAI (GPT-4o + Whisper + Vision)Kommo CRMEvolution APIRedisSupabasePostgres

Agente RAG de AgendamentoRabelo Co. / Rise Doc

● Live

Scheduling and knowledge agent for a salon

Problem: Salon needed 24/7 service and scheduling without blowing up API cost and latency.
Solution: Hybrid RAG: static info served from a vector store, live API reserved only for scheduling. Daily embedding ingestion cron and AI→human handoff. Reverse-engineered an undocumented third-party API (Cognito auth, report endpoints).
Result: Migrated 20 professionals, 224 services and 500 clients to the new schema; cut response latency and cost.

20Professionals migrated

224Services

500Clients

Stack

n8nSupabase pgvectorOpenAI (text-embedding-3-small, 1536d)Cognito/Avec API

CRM multi-tenant com agente de IARise Doc

● In dev

Public chat that qualifies and drops the lead into the panel

Problem: Every clinic needed its own inbound channel, handled with the care health data demands.
Solution: A CRM module inside the platform itself: public chat with an AI agent, lead capture, CFM and LGPD guardrails and consent logging. Per-conversation cost was measured against the API before shipping, to fit the agreed ceiling.

Stack

Next.jsAI SDKgpt-5-miniSupabase (RLS por tenant)

SaaS & platforms

Solid AppSolid Studio

● In dev

Management for 3D printing pros

Problem: 3D printing pros without a tool for store, cost and order management.
Solution: Management PWA: store, catalog, orders, inventory and cost calculation, with Stripe payments and NextAuth v5 auth.

Stack

Next.js 15React 19Prisma 7 + SupabaseStripeNextAuth v5

solidapp.shop

Plataforma de Aprovação e PublicaçãoRise Doc

● Live

The client approves by link and the app publishes on its own

Problem: Content approval over WhatsApp and email: no history, no deadline, and manual publishing on every network.
Solution: Multi-tenant SaaS where each client reviews posts through a public link and, on approval, the app schedules and publishes via the Meta Graph API. Separate agency and client roles, per-project RLS and an auditable approval trail.

Stack

Next.js 16Supabase (RLS)Meta Graph APIVercel

app.risedoc.com.br

AI-powered tools

Face FinderRabelo Co.

● Live

One selfie returns every photo you appear in

Problem: A wedding archive of nearly 4,000 photos on Dropbox: nobody finds their own without scrolling the whole album.
Solution: Archive indexed with face detection and embeddings (ONNX Runtime on GPU), similarity search in pgvector with an HNSW index, and a mobile PWA where the person takes a selfie and downloads a ZIP. The selfie is never stored and consent is asked before the scan.
Result: 3,995 photos and 15,681 faces indexed in ~20 minutes, zero errors, sub-second search. Live in production.

3.995Photos indexed

15.681Faces

<1sSearch

Stack

PythonONNX Runtime (GPU)Supabase pgvector (HNSW)PWAVercelDropbox API

ache-suas-fotos.vercel.app

See every project

In practice

Interactive demo of a real production system — reply as the customer and watch every agent action become a funnel event.

How I build

A real coding-agent session — the one that took this site's LCP from 4.2s down to 1.8s.

rafael@rabelo — claude ~/portfolio-rabelo

Claude Code v2.1.206

Welcome back Rafael!

Claude Opus · xhigh effort

Rabelo Co.

~/portfolio-rabelo

Tips for getting started

Measure before touching

What's new

next/image: priority is out, use loading=eager

Tailwind v4 with no config, tokens in CSS

/release-notes for more

let's optimize my portfolio

Before touching anything: measure. I'll compare against the previous commit in a separate worktree, with throttled network and CPU, so the number isn't a guess.

Measure LCP, FCP and bytes against the previous commit (completed)
Compress the assets (og.png is 1.1 MB) (completed)
Get framer-motion off the critical path (in progress)
Re-validate both ends and ship (pending)

Bash(Fast 3G · CPU 4x · median of 3)LCP 4176 ms · FCP 1924 ms · images 89 KB(ctrl+o to expand)

cause: the root layout mounts framer-motion, so the hero text only paints after hydration

Update(src/components/AnimatedBackground.tsx)

animated background moves from framer-motion to CSS

1-removed: "use client";
3-removed: import { motion, useReducedMotion } from "framer-motion";
4 
5+added: export function AnimatedBackground() {
6-removed:   const reduced = useReducedMotion();
7   return (
8-removed:     <motion.div animate={reduced ? undefined : { x: b.x, y: b.y }}
9+added:     <div className="blob absolute rounded-full"
10       style={{ filter: "blur(90px)" }}

LCP down to 1832 ms (-56%). The hero text now paints on first paint, with no JS in the way — and the background looks identical, animating on the compositor.

Agent interfaces recreated as components (brainless, by theswerd, MIT). The content is the actual session: same file, same diff, same measured numbers.

Companies & Clients

Point at a brand to see what changed in its operation.

Ativos Digitais

Digital marketing, automation and AI for SMBs

3× AI agent + Follow Up · 2× AI agent + Follow Up · 1× Main agent + CRM webhook · 1× AI Sales Dashboard

n8nKommo / PipedriveEvolution APIOpenAI (GPT + Whisper)

CRM (produção) E-book (captura)

What changed

Leads arriving at 3 a.m. started getting answered and qualified on the spot, already logged in the CRM.
Follow-up stopped depending on someone remembering: the stage cadence creates the touchpoints and rolls itself back when the deal moves.
SDR calls turned into context: transcript and summary land as a deal note with nobody typing.
Management started seeing the AI funnel in real time, event by event, instead of asking how the week went.

What changed

Leads arriving at 3 a.m. started getting answered and qualified on the spot, already logged in the CRM.
Follow-up stopped depending on someone remembering: the stage cadence creates the touchpoints and rolls itself back when the deal moves.
SDR calls turned into context: transcript and summary land as a deal note with nobody typing.
Management started seeing the AI funnel in real time, event by event, instead of asking how the week went.

Rise Doc

Digital health — medical automation and AI agents

2× AI agent (SDR + scheduling) · 5× AI agent (SDR + scheduling) · 1× Approval and publishing platform · 1× Multi-tenant CRM with AI agent

n8nOpenAI / ClaudeEvolution APIoAtmos (CRM)

What changed

Booking, attendance confirmation and reminders moved off the front desk and onto the agent — after hours included.
Content approval moved from WhatsApp to a link with an auditable trail, and publishing to the network became automatic.
FAQs and initial triage stopped forming a queue: they're answered instantly, with CFM and LGPD guardrails.

What changed

Booking, attendance confirmation and reminders moved off the front desk and onto the agent — after hours included.
Content approval moved from WhatsApp to a link with an auditable trail, and publishing to the network became automatic.
FAQs and initial triage stopped forming a queue: they're answered instantly, with CFM and LGPD guardrails.

Solid Studio

3D printing and SaaS for makers and pros

Next.js 15React 19Prisma 7 + SupabaseStripe

solidapp.shop

What changed

An archive scattered across dozens of groups became a single search, with AI visually checking the result before delivering it.
Store, catalog, orders, inventory and cost calculation moved into a single app, instead of spreadsheets and notebooks.

What changed

An archive scattered across dozens of groups became a single search, with AI visually checking the result before delivering it.
Store, catalog, orders, inventory and cost calculation moved into a single app, instead of spreadsheets and notebooks.

Elevra

AI virtual assistants for service businesses

1× Virtual assistant with live demo · 1× SDR template (Dify + GHL)

n8nOpenAI (Whisper + GPT-4o-mini)DifyGo High Level

What changed

The first reply to a lead stopped depending on business hours — it lands in seconds, any time.
Customer voice notes and photos stopped being a bottleneck: the agent transcribes and reads them in the same conversation.
The proposal became a demo: the client talks to the agent before deciding, instead of reading a PDF.

What changed

The first reply to a lead stopped depending on business hours — it lands in seconds, any time.
Customer voice notes and photos stopped being a bottleneck: the agent transcribes and reads them in the same conversation.
The proposal became a demo: the client talks to the agent before deciding, instead of reading a PDF.

Unymos

Customer success out of support tickets

1× CS report agent

n8nZoho Desk APIOpenAI

What changed

Ticket volume became a business reading: why the client reached out, what repeats and where the risk is.
The monthly report that used to be compiled by hand now comes out ready, per client.

What changed

Ticket volume became a business reading: why the client reached out, what repeats and where the risk is.
The monthly report that used to be compiled by hand now comes out ready, per client.

VetLíderes

AI service platform for the veterinary sector

1× AI service SaaS platform · 1× WhatsApp channel (addendum)

Vite + React 18TypeScriptSupabasen8n

What changed

The team got a single place for service, instead of conversations scattered across personal apps.
Customer voice notes became searchable text in the professional's history.

What changed

The team got a single place for service, instead of conversations scattered across personal apps.
Customer voice notes became searchable text in the professional's history.

Skills & Tools

The fronts I cover and what I use on each one.

Point at a front to see the tools

Stack & Tools

n8nClaudeOpenAISupabaseNext.jsReactTypeScriptPythonFastAPINestJSn8nClaudeOpenAISupabaseNext.jsReactTypeScriptPythonFastAPINestJS

PostgreSQLPrismaStripeEvolution APIPipedriveKommoClickUpVercelRailwayTailwindPostgreSQLPrismaStripeEvolution APIPipedriveKommoClickUpVercelRailwayTailwind

Open to opportunities

Let's talk

Open to opportunities as an AI Automation Engineer.

Request a proposal Book a call Download CV