Raphael – Python, LLM, RAG

An autonomous software engineering agent that generates, migrates, and updates codebases from natural language specifications. The system maps repositories into semantic knowledge layers, creates structured development plans, and dispatches parallel specialist agents to implement features — replacing manual development cycles with an AI-driven pipeline.

• Rebuilt the semantic compiler to index entire code repositories into structured memory layers, enabling AI agents to reason over codebase context and domain specifications.
• Replaced a generic single-agent pipeline with Mixture-of-Experts (MoE) architecture, routing tasks to specialized agents per domain; improved classification accuracy and output reliability significantly.

A conversational AI assistant embedded in a major Brazilian bank's mobile app, serving retail and investment banking clients. The system handles financial queries, executes trading operations, provides personalized investment guidance, and generates UI components dynamically — backed by a multi-agent orchestration layer with real-time access to cross-institution financial data via Open Finance.

• Architected a multi-agent system using the Orchestrator pattern to route and handle diverse financial intents — trade execution, investment advice, document summarization, and UI generation.
• Delivered a microservice for real-time financial trade operations, abstracting manual complexity for end users.
• Built deterministic internal tooling to constrain agent behavior on high-stakes financial actions, ensuring predictable and compliant outputs.
• Deployed the agent platform on EKS for horizontal scalability and resilience under concurrent user load.
• Implemented a Hybrid Search + Re-ranking RAG pipeline for precise retrieval of financial knowledge base content.
• Integrated the Open Finance MCP to pull cross-institution financial data, expanding the assistant's awareness of the user's full financial picture.
• Optimized RAG for sensitive financial data using Redis Embedding Cache, reducing retrieval latency.

An AI-powered HR automation platform that handles onboarding, ticket creation, and internal requests through a multi-agent conversational interface. The system autonomously generates support tickets from chatbot conversations, schedules meetings by analyzing participant availability, and answers HR policy questions via a RAG knowledge base built from internal documentation.

• Delivered a multi-tenant serverless architecture on AWS, supporting isolated environments for multiple organizations at scale.
• Designed a multi-agent ecosystem in LangGraph with an expert router dispatching tasks to specialized agents for ticket creation, meeting scheduling, and policy Q&A.
• Architected a RAG pipeline ingesting internal policy documents into a structured knowledge base for accurate real-time Q&A.
• Developed an OCR data pipeline using Docling to parse confidential financial documents into machine-readable, structured formats.
• Built the system entrypoint in FastAPI, connected via SQS + SNS event bus to async processing workers.
• Provisioned the full infrastructure as code with Terraform; stored processed files in Amazon S3.

An e-commerce integration platform connecting marketplaces (Shopify, Shein, Shopee) with ERPs and fulfillment providers for Brazilian retailers. The AI layer automates product classification, cross-marketplace schema normalization, and listing quality optimization to boost catalog visibility and seller sales performance.

• Deployed a multi-agent AI ecosystem for digital product classification and quality improvement, directly increasing marketplace listing performance and seller commission rates.
• Rebuilt the cross-marketplace product normalization pipeline from scratch after a breaking schema change (Shein/Shopify/Shopee) two days before a scheduled launch — delivered on time.
• Created graph-based workflows in Neo4j for correlating and correctly mapping products across divergent marketplace schemas.
• Orchestrated autonomous pipelines with LangGraph and Pinecone for high-speed vector retrieval and automated product taxonomy assignment.
• Designed migration agentic pipelines to map legacy codebase entities to new product specifications, enabling schema evolution without manual re-categorization.

A SaaS e-commerce integration platform connecting online stores with ERPs and fulfillment providers for the Brazilian market. The platform processes high-throughput order and product events through an event-driven architecture, enabling reliable multi-channel operations across major Brazilian marketplaces.

• Architected 15+ e-commerce integrations using an event-driven approach with RabbitMQ and Redis, processing high-volume order and catalog events across marketplaces.
• Deployed integrations on Kubernetes for container orchestration and service reliability.
• Built logistics tooling for e-commerce operations, streamlining fulfillment and shipping workflows.
• Integrated the platform with core AWS services (Lambda, S3, SQS, SNS) for event processing and async storage pipelines.