Matheus – AI agent orchestration, Python, LLM

An internal AI platform that enables product teams to deliver generative AI capabilities across a SaaS ecosystem for home-service businesses. The platform provides orchestration, deployment, and observability infrastructure for agent-based and LLM-powered features, supporting reliable performance, low-latency execution, and scalable AI adoption across products.

• Built and evolved the internal AI platform powering generative AI across multiple products;
• Designed and implemented agent-based systems for task automation and developer workflows, including orchestration layers for tool use, execution, and long-running AI tasks;
• Improved reliability, latency, and cost efficiency of LLM-powered systems in production;
• Implemented observability and feedback loops for continuous improvement.

An AI-native learning platform designed to help users transition into technology careers through personalized, adaptive learning journeys. The product uses multi-turn LLM interactions to assess skills, generate individualized learning roadmaps, validate knowledge through AI-generated assessments, and continuously adapt recommendations based on identified learning gaps. Core product functionality is driven by AI-powered reasoning, planning, evaluation, and remediation workflows.

• Architected and shipped the full AI execution layer;
• Designed and implemented six LangGraph workflows covering diagnosis interviews, roadmap generation, validation, mock interviews, and mentoring experiences;
• Built evaluation loops and decision frameworks to support adaptive learning plans, assessment quality, and workflow reliability;
• Owned the delivery harness, including a triple QA gate with automated evaluation, browser-based validation, and structural verification;
• Coordinated parallel task execution for priority Linear issues and managed the development lifecycle through Linear-based workflows;
• Maintained a self-rebuilding STATUS.md parity matrix to ensure delivery and documentation consistency.

An autonomous AI agent designed to manage software delivery workflows from ticket intake to implementation. Inspired by multi-agent orchestration systems, the platform monitors Jira, identifies tagged work items, and advances them through a structured development lifecycle using deterministic state transitions, rubric-based evaluation, and human approval checkpoints.

• Contributed to the development of an autonomous agent platform for software delivery automation;
• Implemented human-in-the-loop approval checkpoints using Slack emoji reactions;
• Built workflows that dispatch Claude-driven development tasks within sandboxed execution environments;
• Enabled agents to use pre-warmed Python environments for code generation, execution, and testing;
• Contributed to a hackathon project that won 1st place.

An AI-powered analytics platform for revenue and commissions reporting within a large investment banking environment. The platform enables business users to interact with complex financial data through natural language, combining business-rule-aware Text-to-SQL, agent-driven analytics workflows, streaming responses, and voice interfaces to accelerate reporting and decision-making.

• Architected and owned the agent runtime, with agents defined declaratively in Postgres and compiled at query time into LangGraph execution graphs served through FastAPI with streaming and voice capabilities;
• Built a Model Context Protocol (MCP) ecosystem from scratch, creating an orchestration layer that connected to internal retrieval services providing RAG and GraphRAG capabilities for business-rule-aware Text-to-SQL;
• Shipped an offline evaluation pipeline with release guardrails that blocked agent versions regressing on quality, latency, or cost;
• Built a data flywheel for continuous improvement by streaming Langfuse and Postgres traces to S3 and assembling supervised fine-tuning and preference datasets from labeled agent trajectories;
• Developed an event-driven distributed enrichment platform on AWS Step Functions and Lambda using Observer and Pub-Sub patterns for high-throughput, low-latency data processing;
• Added automated chart generation and presentation export as agent tools, reducing reporting turnaround time from hours to minutes.