Gustavo Karsten

In 30 seconds

I'm an AI agent engineer and CTO & Co-Founder of DockPlus AI since January 2026. I build agents that act on a real business's data, MCP servers, data pipelines and the integrations around them, for clients in Brazil and the US.

I've worked in IT since 2005. Most of that time was in Linux infrastructure, so I know how to put a system into production and keep it running. Since 2024 I've been working with LLMs, RAG and agents.

I'm based in Santa Catarina, Brazil, and work with US clients in English.

What I build

• AI agents with tools: a custom agent loop, human confirmation for irreversible actions, permissions inherited from the user.
• MCP servers: the product's own tools exposed to Claude and other clients, with per-user keys and audit logs.
• Data pipelines: collectors that keep running when a source breaks, with raw data stored so it can be reprocessed.
• RAG: hybrid search, reranking and citations that point back to the source.
• Multi-tenant SaaS: tenant isolation, Stripe billing, integrations with external APIs.

Featured projects

LedgerJob — ledgerjob.com

SaaS for US construction contractors. Owners manage leads, quotes, projects and invoices by chatting with an agent. I designed the backend, the agent and the MCP server: 66 tools back both the chat and MCP, and 16 of them require human confirmation. Integrates Stripe, QuickBooks Online, Twilio and Google Calendar.

Cape Cod Permits

Pulls building permits from 74 Massachusetts towns, spread across 6 platforms with no shared API, into one PostgreSQL database. People and agents query it through a dashboard and a read-only MCP server with 7 tools. Each portal has its own collector, and a failure in one portal doesn't stop the others.

Achado de Leilão — achadodeleilao.com.br

A single search across official Brazilian auctions, live since Sept 28, 2026. A Python collector covering 22 official sources, a Next.js site and an MCP server with 7 tools, including read-only SQL with a timeout and a row cap.

DockPlus Atendimento

A customer service agent for WhatsApp, Telegram and email. A LangGraph graph classifies intent and chooses between a RAG answer, scheduling or handoff to a person. WhatsApp runs on Meta's official Cloud API, webhooks go through a Postgres queue so a message is never sent twice, and the admin panel has RBAC and MFA.

Reachant — reachant.com

B2B lead-list builder for Brazil and the US that enriches every lead. Credits are debited without races between workers, the Stripe webhook is idempotent, and scraping is protected against SSRF. Includes an API and an MCP server with 9 tools.

NutriApp — usenutriapp.com

Practice software for dietitians, with each practitioner's data isolated by PostgreSQL Row Level Security. AI reads lab reports, with mandatory human review. Clinical consultant: Dr. Cilana Carvalho.

DockPlus AI client websites — dockplusai.com/portfolio

I built the sites for Teqbio, Castle Masonry, Cerâmica Santo Amaro, A.F Consultoria, NotaLeve, Larissa Ovalhe and Instituto Odonto Excelência. The process became a Claude Code skill the team uses: research, brand book, site and deploy, with a human approval step between phases.

AgentLab — github.com/gustavo-karsten/AgentLab

An open-source study lab where I test agent architectures. A 5-node LangGraph graph with every step visible live, Langfuse tracing, and user API keys encrypted with AES-256-GCM.

Video lesson RAG

A course platform where students ask questions about a lesson and get answers with citations that jump the video to the right moment. HyDE, hybrid search (vector + full-text) and reranking, with queued async ingestion.

IRS Helper

RAG over IRS forms and publications. It crawls documents from irs.gov, indexes them in pgvector and answers through an OpenAI-compatible API that OpenWebUI uses as if it were a model.

Stack

How I work

• I build with Claude Code and Codex. They're part of my process: I set the architecture, the rules and the tests, and I review what they write.
• Spec before code. AgentLab was built from 12 per-feature specs.
• Business rules in the right place. If an agent can call a tool, validation lives in the service layer, not in the HTTP route.
• From requirements to monitoring. I scope the problem with the client, build it, deploy it and watch it in production.

Infrastructure background

Before AI, I worked with Linux and Windows servers, high-availability databases, firewalls and Ansible provisioning. That background helps me ship and run what I build.

Experience

• DockPlus AI: deliver AI systems from requirements to deploy, using LangGraph, pgvector RAG and Langfuse to track latency, token costs and response quality.
• PowerX: built LangChain and Ollama pipelines connected to internal data and APIs; document-grounded RAG with pgvector. Ran high-availability Linux and database infrastructure, with Ansible provisioning.
• Grupo Fleury: supported critical healthcare infrastructure, Active Directory and ServiceNow.
• Toor Tecnologia: Linux/Windows servers, firewalls, Hyper-V and automation scripts.

Education, certifications and languages

• Autonomous Agents and Knowledge Automation — Estácio de Sá University (in progress).
• Certifications: Linux Fundamentals (FIAP); Technical Leadership (ApoioAG).
• Languages: Portuguese (native); English (upper-intermediate, working with US clients); Spanish (intermediate).

Contact

• Email: [email protected]
• LinkedIn: linkedin.com/in/gustavo-karsten-88a04383
• GitHub: github.com/gustavo-karsten