Qualify and route
Claude returns structured lead qualification while LangGraph makes the workflow explicit and auditable.
Featured project · GTM AgentOS
I build AI systems that turn complex operations into clear, auditable action.
Full-stack AI engineering across agents, automation, retrieval, controlled tools, and human approval — designed to make every important decision reviewable.
Available for AI product engineering rolesAI agentsRAGControlled toolsHuman-in-the-loop
GTM AgentOS / Flagship project
Featured project · GTM AgentOS
AI-native Revenue Operations, with evidence and control built in.
A production-style platform that qualifies leads, orchestrates agent workflows, grounds research in internal knowledge, and gates sensitive actions behind human approval.
Read the engineering case studyGround the research
Voyage embeddings and PostgreSQL with pgvector retrieve relevant GTM knowledge before generation.
Control every action
MCP allowlisting, HMAC-signed n8n workflows, idempotency, and approval gates constrain execution.
118automated tests
E2Emajor flows validated
HITLsensitive action approval
Audittraceable state and tools
Production-style engineering — not presented as a production deployment.
See the system in 84 seconds
From a HOT lead to grounded, auditable action.
The demo follows qualification, LangGraph routing, RAG evidence, controlled MCP tools, human approval, and the complete lead timeline using synthetic portfolio data.
Portfolio Mode · synthetic data · no provider called during capture
Engineering depth
Built around boundaries, evidence, and observability.
The strongest work is not the model call. It is everything that makes the call reliable, constrained, and explainable.
01
Structured LLM outputs keep qualification inside a validated contract.
02
Grounded retrieval includes a safe no-context fallback instead of invented evidence.
03
Controlled execution combines MCP allowlists, HMAC verification, replay protection, and idempotency.
04
Production-style observability records state transitions, tool calls, latency, and AI usage without exposing chain-of-thought.
PythonFastAPISupabase / PostgreSQLClaudeLangGraphVoyage AIpgvectorMCPn8n
Selected work · not a complete list
Two more systems, presented with the work behind them.
This portfolio highlights a curated selection of projects from a broader body of work. Each one shows a different operational problem, product boundary, and engineering approach.
PROJECT 02 · FINANCE OPERATIONS
LedgerGuard AI
A finance operations system that reconciles payouts deterministically, explains exceptions with bounded AI, and keeps consequential decisions human-reviewed.
LedgerGuard AI · complete investigation82-second product demo
Problem
Prove why an expected payout and the received amount do not match.
System
Deterministic reconciliation first, bounded AI explanation second.
Control
Human review remains mandatory before consequential action.
Next.jsTypeScriptSupabaseClauden8n
PROJECT 03 · RECOVERY AUTOMATION
Hotmart Recovery Engine
A portfolio-safe recovery system with a verified 22-node n8n workflow, three state-aware follow-up stages, and an observable operations dashboard.
Hotmart Recovery Engine · portfolio-safe workflow52-second product demo
Workflow
22 n8n nodes evaluate eligible purchase journeys every five minutes.
Recovery
Three state-aware stages select the next follow-up without repeating earlier sends.
Visibility
Supabase history and a read-only dashboard keep every stage observable.
n8nSupabaseWebhooksState machineQA automation
Selected portfolio projects — this is not a complete list of my work.
About Igor
I build the product, the automation, and the proof that it works.
I'm Igor Marques, an AI Product Engineer based in Brazil. I combine full-stack development, operational automation, and product thinking to turn ambiguous workflows into clear, testable systems.
My focus is practical AI: deterministic logic where correctness matters, models where interpretation helps, and humans at the final decision point.
Let's build something useful