Skip to main content

Compliance-grade AI architecture · RAG · agents · MCP

Most AI stacks fail audit at the same three seams — prompt provenance, retrieval lineage, output attestation.

I architect RAG and agent systems for regulated workloads — banking, tax, legal — where the audit trail is the deliverable. Every figure cites its source. Every decision reconstructs six months later. The cost ceiling holds at peak. Built to LGPD, BCB 4.893, and the EU AI Act’s Article 12 logging mandate.

0
hallucinated numerical fields
18 months · regulated tax-filing agent in production
0K+
daily active users
multilingual RAG · +40% knowledge-base precision
0K+/day
transactions cleared
event-driven retail backend · outbox + idempotency
0%
less manual review
tier-1 LATAM legal-domain RAG · validated citations
fig. 01selected work

Selected work

The constraint that made it hard, the decision taken, and the measured outcome — not just the result.

fig. 02open-source & demos

Open-source & demos

fig. 05how an engagement scales

Start with a verdict. Scale when the proof holds.

Each offer ships a defined deliverable, scoped and fixed at signature. Most clients open with a diagnostic, then expand into a build once the architecture proves out against their own data.

Patternprove it

Regulated-AI Architecture Review

A fixed-week diagnostic of your existing LLM stack — three-plane topology memo, severity-ranked findings, annotated reference repo.

Outcome — A go/no-go verdict and a scoped remediation map before you commit engineering quarters.

MCP Tool-Boundary Security Audit

STRIDE threat model of every exposed tool, LLM-vs-operator input-boundary review, deny-by-default permission matrix, signed-release hardening (Sigstore + SLSA L2 + dual SBOM).

Outcome — A severity-ranked report with concrete patches, and a pipeline where every binary verifies with one command.

Pipelinemake it real

RAG Audit-Chain Readiness Sprint

A production retrieval pipeline — pgvector + hybrid retrieval + rerank, forced-citation answers, recall measured on your own holdout set, a decision-trace ledger keyed to (prompt, docs, model, output).

Outcome — Provably-grounded answers, accuracy measured against your data, and behaviour auditable from day one.

Event-Driven Backend Build & Rescue

An authenticated, production-shape backend — typed schema, audit ledger, outbox + idempotency, fitness-function tests, CI gate, observability. Serverless variant ships at $0 idle.

Outcome — A backend that survives load, costs nothing idle, and provisions and tears down reproducibly — owned in your repo.

Platformkeep custody

Embedded AI-Platform Custody

Fractional architecture custody — weekly fitness-function review, monthly audit-chain integrity probe, compliance-plane ownership, participation in the AI hiring loop.

Outcome — An audit-grade AI capability your whole org reuses, with the audit chain kept green between releases.

Every engagement opens with a short discovery call and a written diagnostic. Scope is fixed at signature. Send a brief →

Technical Skills

Grouped by the problem each stack solves, so you can scan for the one that matches yours.

Regulated AI & compliance

RAG with retrieval lineage, hash-chained audit ledgers, and decision provenance — mapped to LGPD, BCB 4.893, and the EU AI Act Art. 12 logging mandate.

LGPDBCB 4.893EU AI Act Art. 12Audit-trail designDecision provenanceRAG evaluationPII controls

AI agents & RAG

Production RAG pipelines, typed-tool agent loops with bounded turns, and Model Context Protocol integrations — grounded retrieval with frozen-eval regression checks.

Claude CodeAnthropic SDKAzure OpenAIAWS BedrockRAGMCPpgvectorLangGraph

Backend

Polyglot services that hold under load: Go daemons, TypeScript APIs, Python pipelines, Rust binaries — event-driven with outbox + idempotency.

GoTypeScriptNode.jsPythonFastAPIRustPostgreSQLKafka

Cloud

Multi-cloud deployments on AWS, Azure, and GCP — serverless and container workloads sized to a cost ceiling, not left to drift.

AWS LambdaAWS BedrockAzure App ServiceAzure AKSAzure OpenAIGCP Cloud RunGCP Cloud SQL

Infra-as-Code

Declarative infrastructure across cloud and bare-metal fleets — reproducible, with a teardown story that leaves zero orphans.

TerraformHelmKubernetesNixOSDockerDocker Compose

Release engineering

Supply-chain hardening as a first-class deliverable: reproducible builds, signed provenance, dual-format SBOMs.

GitHub ActionsSigstoreSLSA L2gitleaksOSV-ScannerDependabotSyft

Experience

Shipped systems and the outcomes they moved.

Compliance-Grade AI Architect / Cloud Architect

Jul 2025 — Present

Tier-1 IT services group · LATAM

  • Architected compliance-grade RAG on Azure OpenAI — decision provenance, audit-trail logging, and frozen-eval regression checks for regulated workloads
  • Shipped a multilingual education assistant serving 100K+ daily active users — +40% knowledge-base precision after rollout
  • Stood up multi-cloud Terraform (Azure + GCP) cutting environment provisioning below 10 minutes
  • Cut cloud spend 30% via Lambda right-sizing and reserved-capacity planning

Systems Software Engineer

Oct 2024 — Sep 2025

Telecom carrier · LATAM

  • Designed serverless ETL on AWS Lambda + Step Functions for tier-1 telecom billing data
  • Published Terraform modules provisioning multi-region infrastructure in under 10 minutes
  • Rolled out a CloudWatch observability stack — dashboards, alarms, automated incident response
  • Hardened the release pipeline with Sigstore + SLSA provenance for a regulated supply chain

Senior Software Engineer

Jul 2021 — Oct 2024

Product engineering studio · e-commerce / fintech / logistics

  • Delivered 12+ production systems across e-commerce, fintech, and logistics
  • Built an event-driven BFF + Broker + Dispatcher clearing 10K+ transactions/day with the outbox pattern and idempotency keys
  • Introduced GitHub Actions matrix CI with gitleaks + OSV-Scanner for supply-chain hygiene
  • Set the architecture and release-gate standards adopted across multiple squads
fig. 06claim → evidence

Every claim here is auditable

Six public repositories under yolo-labz, each SLSA L2, Sigstore-signed, and gated on live SonarQube quality checks. Read the source, not a screenshot. Client work links to anonymized writeups, never names.

Contact

I build for the engineer who gets paged at 02:00 BRT and needs the audit chain to still hold. The cost ceiling stays put at peak. The writeup still reconstructs six months later. Send a brief on architecture, RAG, compliance, or supply-chain. I'll tell you straight whether it's a fit.