Computer Engineering student (UNIFRAN / Cruzeiro do Sul Virtual) and Economics graduate (UFSJ). Based in Barbacena, Brazil. I build small tools with Python and write down what breaks on the way, because most of what I know came from debugging something I broke myself.
Currently looking for a remote internship in backend Python / AI tooling.
A pure-Python coding agent with measured context caching and a containment layer for command execution, evolving from blocklist to allowlist to origin validation to surface reduction to isolation. Context caching cut cost by 76.2% in a real benchmark; the execution layer went through ten rounds of adversarial review by a multi-model panel, and the sandbox work is next.
Python · 140 tests, CI green
→ github.com/brmarcosbr/harness
A mod plus a Python/MCP bridge that feeds the game state to an agent and returns one command per turn. The autopilot is a deterministic greedy policy with no LLM in the loop, and the game's refusals are fed back as state so the bot stops repeating an illegal move.
Python · C# · MCP
→ github.com/brmarcosbr/sts2-mcp-bridge
→ write-up: "I made an agent play Slay the Spire 2 on its own"
A Vite + TypeScript app that solves, generates and scans Sudoku grids: step-by-step logical hints (naked/hidden pairs and triples), unique-solution generator, and a scanner that reads printed puzzles with a local CNN plus an LLM fallback. 170+ tests.
TypeScript · deployed · private repo
Python (FastAPI) · TypeScript · C · Linux · Docker · Git · MCP · a bit of
C#/Harmony for game mods
I prefer a number over an adjective: if I claim something is faster or cheaper, there is a measurement behind it, and a log of the attempt that failed first.
✍️ dev.to