Senior Generative AI Engineer focused on enterprise AI platforms and the trust infrastructure required to operate AI systems safely: retrieval, agents, identity, authorization, observability, runtime policy enforcement, evaluation, and verifiable governance.

My work combines software architecture, LLM engineering, security, LLMOps, and governance for production systems in regulated environments.

I bring more than 22 years of experience across financial services and technology, including Caixa, BTG Pactual, Banco do Brasil, Itaú Unibanco, and ASA SCFI.

Explore the complete portfolio architecture and project relationships

The repositories form an engineering ecosystem rather than a collection of isolated demos.

The current architecture explores four complementary layers:

• Governance and assurance: policy, risk, approvals, evidence, runtime assurance, incident response, and auditability.
• Runtime trust services: model-policy enforcement, OAuth/OIDC for remote MCP, A2A/MCP trace continuity, and privacy-safe telemetry.
• Domain AI systems: RAG, enterprise knowledge, Open Finance tools, and auditable multi-agent credit analysis.
• AI-assisted engineering controls: deterministic execution, canonical evidence, architecture constraints, and human-controlled promotion.

The common design goal is to keep high-impact decisions and security boundaries outside model reasoning while still using LLMs where generative behavior adds value.

• Built and governed production AI systems for regulated financial institutions, including conversational assistants, RAG pipelines, agent workflows, observability, and compliance controls.
• Led enterprise AI adoption for approximately 400 users, including Claude Code for around 250 developers and Claude Enterprise for approximately 150 business users.
• Reduced the average context of an investment assistant from approximately 70,000 to 3,000 tokens through conditional knowledge injection, improving accuracy while reducing latency and inference cost.
• Established an AI governance function covering usage policies, approval processes, MCP allowlists, risk assessment, auditability, incident response, and phased enterprise adoption.
• Translated governance requirements into a verifiable implementation with deterministic controls, segregation of duties, evidence-bound decisions, runtime enforcement, and tamper-evident audit history.

Alicerce · engineering-loop-schemas · Claude Python Engineering Harness · Codex Python Engineering Harness

Open Finance BR MCP · Meridian

• Critical business decisions remain deterministic and auditable.
• Authentication and authorization are enforced in code, never delegated to a language model.
• Agents, MCP servers, model providers, retrieved data, and telemetry pipelines are treated as trust boundaries.
• Structured outputs, explicit contracts, bounded retries, and fail-closed validation constrain generative behavior.
• Evidence must be independently verifiable; model self-reports are not proof.
• Approvals remain bound to the exact reviewed scope, and material changes trigger reassessment.
• Telemetry is minimized by design; prompts, model responses, credentials, and business payloads are not observability defaults.
• Promotion, deployment, runtime overrides, and other high-impact actions remain under explicit human authority.

Generative AI Engineering
RAG · agentic systems · multi-agent architectures · semantic and model routing · structured outputs · tool calling · LLM evaluation · guardrails

Agent and MCP Security MCP · A2A · OAuth 2.1 · OIDC · Microsoft Entra ID · PKCE · resource/audience binding · least privilege · fail-closed authorization · prompt-injection boundaries

AI Platforms and Governance Enterprise AI platforms · policy enforcement · runtime authorization · model and agent assurance · evidence · auditability · incident response · human-in-the-loop

LLMOps and Observability OpenTelemetry · W3C Trace Context · OTLP · Datadog · Langfuse · distributed tracing · structured logging · latency percentiles · evaluation pipelines · regression testing · token/cost observability