verified offline, live smoke pending
SupportOps
A support platform where AI can draft a refund, but only a second person can approve it.
Diego Vieira Gonçalves de SouzaAI Engineer
AI Engineer building agents you can verify.
Four years of software engineering, two in generative AI, shipping production agents. This site runs a live multi-agent support platform. Trigger it, watch every step, then read the code and the evals behind it.
Open the live demoRead the flagship case
What you can check here
27 of 36
Real-model trials passed (development set)
Development set, not held out. Status FAIL_PROVISIONAL; the exit code follows that status, as the P00 README defines it.
17 passed
demo-live-v1 ledger and failover tests against PostgreSQL 16
Fake provider transport. No model was contacted.
2 passed
Requester cannot approve their own refund; a tenant cannot read another tenant's ticket
One ticket, every project
The projects are not separate demos. They are the parts of one system. Press play, or open a node to see the code at its exact commit.
The whole flow. Press play or next to walk through it.
Text version of this diagram
- Ticket P06. A visitor opens a ticket in the console.
- then API and auth
- API and auth P05. Tenant and role come from the signed token, not from the request.
- then Guard screen
- Guard screen P02. Text is screened before any model sees it.
- then Durable job
- Durable job P04. The work is a leased job with retries and a dead-letter path.
- then Four agents
- Four agents P01. A supervisor runs classifier, researcher, billing and responder, with a hard step limit.
- then Real model
- then Content-free trace (spans)
- Real model P00. Each call is reserved in a ledger before the provider is contacted.
- then Refund gate (proposal)
- Refund gate P05. A deterministic check turns a proposal into an approval request.
- then Second person
- Second person P06. Someone other than the requester approves or rejects.
- Content-free trace P03. Spans carry timing and status, never the text.
Flagship
Projects
- failed provisionally, reported
Foundation and live evals
The contracts every project shares, and the evaluation that told me my agent was not ready.
- offline verified, quality pending
Multi-agent orchestration
Multi-agent triage with a bounded loop, compared across frameworks on the same contract.
- fixtures verified, quality not claimed
Production AI toolkit
The boring safety layer that keeps an agent from hurting production.
- service gate verified, production telemetry pending
Agent observability
See what an agent did without storing what users said.
- isolated gate verified, no cloud run
Durable async workflows
Agent work that survives a worker dying mid-task.
- local gates pass, not published yet
This site and its design system
A design system with contrast checked by tests, and a site that refuses unproven numbers.
- process in use, outcome per project
The engineering harness
The engineering system that built all of the above with coding agents.
How I build with AI
Built with coding agents under an engineering harness: specs before code, gated checks, evals, budget ledgers and saved evidence, following AI engineering best practices. Every number here links to the run that produced it.
Experience
4 years of software engineering, 2+ years with generative AI, 1+ year in production. I build AI agents and integrate Claude and OpenAI models into systems through REST APIs, webhooks and tool use, with Python, RAG on pgvector, LangChain, LangGraph and MCP, on AWS.
- AI Engineer, Grupo Intra de Ensino e Pesquisa a Distância. 2025-03 to 2026-05 · Career history, not a measured result.
- Software Engineer, Yamaha Motor (via Compass UOL). 2024-02 to 2025-01 · Career history, not a measured result.
Contact
Remote, from Crato, CE, Brazil. WhatsApp is the fastest way to reach me; email, LinkedIn and GitHub also work.