Production AI agents and integration pipelines — built, hardened, and kept running.
I build LLM agents and API integration pipelines that run in production, against real money and real customers. The interesting work is rarely connecting the APIs — it is containing what the model gets wrong when nobody is watching.
At Escalys, I architect and maintain a REST/webhook integration mesh (Pipefy → n8n → Asaas → WhatsApp) supporting an operation with 11 pipelines and 4 databases, plus Python AI agents that classify documents and extract structured data from PDFs.
Independently, I run a WhatsApp ordering agent in production — OpenAI tool-calling, n8n orchestration, Redis state — handling ~100 real orders every two days through a third-party REST API.
Background spans IT support and infrastructure before development, so I own a workflow end to end: design, integration, failure testing, root-cause analysis and documentation.
Open to remote work — full-time or contract. Registered business entity, invoices in USD. Professional working proficiency in English.
These three read in sequence: the incident, the pattern it produced, and the tool that generalises it.
|
Case study of a production AI ordering agent — the failures that only appear with real customers, and the deterministic guards that contained them. Includes the validation node's code.
|
Sixteen patterns for running AI agents in n8n against real money, each traced to a specific incident — including the parts of the n8n execution model that are not documented anywhere.
|
FastAPI service that validates an agent's tool call before it becomes an irreversible action. 44 tests, CI across Python 3.11–3.13, Docker image.
|
- Root-caused a silent production failure that had run for months — an order-status lookup written off as vendor flakiness. One unwired branch, raising no error. Found by auditing connections against expected behaviour, not by guessing.
- Root-caused an agent emitting invalid product IDs — the vendor API exposed an undocumented internal field sharing the numeric shape of real IDs. Two prompt-level fixes had already failed; the fix was a deterministic code guard.
- Reverse-engineered undocumented vendor API behaviour — combo pricing rules and fee resolution, from response evidence alone.
- Tested credential and network failures against real endpoints on an isolated clone, without ever risking live service.
| Process Automation & AI Integrations Engineer — Escalys | Jun 2026 – Present |
| IT Generalist / Technical Support — Casa Clínica | 2023 – 2025 |
| Full Stack & Automation Developer — Freelance | 2022 – Present |
Technologist in Information Technology Management — UNIAENE-FADBA (2023–2025) · Databases (Oracle) · Machine Learning (Microsoft)
Client automations built at Escalys — technical-report extraction in Pipefy, transactional email in n8n — have no public repository. The full case studies are on the portfolio.
AI & Automation
Backend
Data
Infra & DevOps
Frontend
| Repository | What it is | Stack |
|---|---|---|
| atend-ia | Chatbot with context persistence, built for integration with support channels | FastAPI · Claude API · PostgreSQL · Redis |
| api-tarefas-spring | REST API with DTOs, validation, filters and global error handling | Java 17 · Spring Boot 3.2 |
| ai-rag-dashboard | RAG over your own documents — upload, ask, get answers with their sources | FastAPI · React · OpenAI · pgvector |