• Architect, prototype, and deploy RAG pipelines, combining vector search, hybrid retrieval, reranking and contextual compression techniques. • Contribute to design and orchestration of multi-agent LLM systems using community frameworks and custom orchestration layers. • Work on a variety of information extraction, information storage and information retrieval problems for both structured and unstructured data. • Partner with cross-functional (product, infra, data engineering, and software engineering) to build robust, high-scale systems that underlie all of our data processing and ML Operations.

• Build and operate the LLM and agent infrastructure every team at Coinbase builds on. • Ship high impact AI agents on top of the LLM infrastructure. • Serve as the company's single path to LLMs and the substrate for the full agent lifecycle: build, deploy, run, observe, improve.

• Você vai construir a plataforma de agentes de IA da Guanabara e o ecossistema que distribui IA generativa governada à companhia. • Orquestração de agentes — construir agentes com LangGraph no padrão Supervisor/ReAct, com memória de curto/longo prazo e estado durável (PostgresSaver). • Camada semântica e RAG — evoluir a camada semântica determinística (Cube.js) e a recuperação semântica com Qdrant (busca vetorial + re-ranking) que ancoram as respostas em dado real. • Skills, subagentes e MCP — criar e versionar skills (SKILL.md), subagentes e ferramentas MCP (FastMCP) para o Claude Code, em fluxo spec-driven, conectando agentes a sistemas corporativos. • Verificação e avaliação — manter o verificador (LLM-as-judge) e suítes de eval (offline/online) com amostragem de traces para pegar drift de qualidade. • Plataforma de modelos — operar o gateway LiteLLM (Virtual Keys, budget, RBAC, roteamento multi-provider) otimizando custo e latência. • Adaptação e serving de modelos — quando RAG e prompting não bastam, fazer fine-tuning supervisionado e PEFT (LoRA/QLoRA), quantização e destilação; servir modelos abertos/privados com vLLM e roteamento SLM↔LLM. • Governança e segurança — aplicar guardrails de PII (Presidio), defesa contra prompt injection e observabilidade ponta a ponta (Langfuse).

• Serving as Scrum Master and Delivery Lead for both AI teams: organizing and facilitating sprint planning, daily stand-ups, backlog grooming, and retrospectives. • Shielding both teams from day-to-day integration distractions by ensuring the junior development team receives clean task definitions, structured schemas, and clearly scoped technical requirements. • Balancing high-speed AI prototyping demands against the structured pipeline stabilization cycles required for enterprise-grade development. • Managing cross-team dependency and interface mapping to ensure smooth collaboration between the senior and junior engineering layers. • Translating strict architectural guardrails — network isolation, database connection limits, cost-containment — from the System Architects into practical workflows for the engineering teams. • Partnering with Loftware Architects to ensure teams safely leverage AWS services and data read replicas without compromising corporate security boundaries, tenant isolation, or regional compliance. • Leading technical review sessions to determine the appropriate storage strategy (Amazon MemoryDB / Redis OSS / Valkey vs. pgvector vs. OpenSearch), balancing developer needs against enterprise infrastructure standards. • Overseeing evaluation frameworks for multi-step agent workflows to ensure deterministic behavior and eliminate unhandled hallucinations. • Validating that all data ingestion flows and internal tool-calling structures adhere to type-safe validation layers, preventing malformed agent responses from breaking downstream systems or leaking PII. • Overseeing the centralized repository for system prompts, prompt caching strategies, and Amazon Bedrock configurations to ensure optimal performance, token budgeting, and corporate policy alignment. • Working with internal teams to define and enforce robust CI/CD strategies for AI agents, ensuring that changes to prompts, embeddings, or state-machine routing rules are deployed without service disruption. • Contributing to operational protocols for deployment failures mid-workflow, ensuring both teams design for idempotency to handle unexpected model degradation or pipeline failures gracefully.