Backend developer from Salvador, Brazil. I run MoWave, the software studio behind Lima, an AI agent that acts on your calendar, tasks, habits and money instead of answering questions about them. Launching October 2026.

Most of my days are Java and Spring Boot. Python is where my open source lives, and Flutter is what puts the app on phones. What actually holds my attention is agent design, and lately the memory side of it. How a model picks the right tool, what it should remember about you, and when it should just do the thing instead of asking.

loci is a memory layer for AI agents. pip install mowave-loci

An agent that forgets you every session is a search box with manners. Loci keeps durable facts about a person as typed records, pulled out of the conversation after each turn instead of typed into a settings page. A daily job merges duplicates and expires guesses nobody confirmed. Only the twelve most important facts get injected, so the context budget stays for the actual conversation.

The part I did not expect going in is that a memory layer is a prompt injection surface. Whatever gets stored lands in every future system prompt, forever, so loci cleans stored text before it goes in.

The core has no dependencies at all. On top of it sit an MCP server that works in any client and a Hermes plugin that adds automatic extraction through post_llm_call. MCP cannot do that part, since nothing calls a server when a turn ends.

loci-decisions-agent is the agent that proves it works. Once a day it asks loci what is due for review, sends one question per fact over WhatsApp, and updates the store from the answer. With no memory it does nothing at all, which is the whole point.

Lima keeps your calendar, tasks, habits, focus and money in one place, with an agent on top that can actually act on all of it. Ask about a flight and it checks whether the date is clear and whether the money is really there before answering.

A few pieces I enjoyed building.

• 65 tools behind native function calling, with a bounded loop and parallel execution inside one iteration
• Specialized subagents behind a keyword classifier that routes in under 5ms with no model cost, falling back silently to the generalist when one of them fails
• Open Finance through Pluggy across 200+ banks, so the agent reasons over real balances instead of numbers you typed in
• Voice, image and document input on one engine, with no separate transcription or OCR service

Java 21 and Spring Boot on the backend, Flutter on mobile. WhatsApp is built and waiting on platform approval.

Cassandra runs Monte Carlo and Markov chain simulations that map where a stock price can go, with backtesting and risk controls. Python

Penumbra is an open data index of the most overlooked municipalities in Bahia, plotted on an interactive map of the cocoa region. Python

reconhecimento_sono is computer vision that notices when someone starts falling asleep and fires an alarm. Python

tomorrowUFBA won UFBA's programming competition. I led a team of eight on it. HTML

Most of Lima lives in a private repo, so what is public here is only part of the picture.