Fabio Nonato
de Paula
Simple systems that reason and act
Principal Applied Scientist @ AWS
Simple systems that reason and act
Principal Applied Scientist @ AWS
I build AI systems that reason, remember, and act — with a bias toward simplicity and local-first design. At AWS, I work on the Agentic AI org. Outside of work, I build custom agents, MCP servers, developer tools, and the occasional game — all guided by the principle that if you can't audit it, you shouldn't ship it.
Custom minimal AI agent with persistent memory, MCP tools, and Discord — zero frameworks
Local FAISS vector store as an MCP server — drop-in semantic search for Claude / Copilot / Agents
Terminal UI for browsing, searching, and resuming Claude Code sessions
A quiet place to read Markdown — native macOS e-reader with highlighting, annotations, and GFM support
Live transcription with Whisper — microphone and system audio capture
Grimdark twin-stick shooter with procedural arenas and adaptive AI director
Building AI agents that reason, plan, and act autonomously — from custom frameworks to production systems at AWS.
MCP servers, FAISS vector stores, graph retrieval. Local-first tooling for semantic search and agent memory.
eBPF-based guardrails, threat intelligence, and securing AI systems — from kernel-level monitoring to policy enforcement.
TUIs, hardware drivers, and productivity tooling for AI-augmented workflows. Building the tools that build the tools.
I rebuilt my Pygame shooter in Godot 4 with an agent. The 95k lines of code were the easy part. Keeping four pictures of the same orc consistent was not.
Software was never as predictable as we told ourselves. AI didn't break the promise. It exposed that the promise was always fiction.
Why I rejected every agent framework and built Luna — a ~2,300-line Python AI agent with SQLite hybrid-search memory, MCP tools, and Discord, fully local.