Computer engineer · Curitiba, Brazil
About
Computer engineering at PUCPR, in Curitiba, Brazil. Two companies of my own, two contracts, and a stubborn habit of measuring things.
I started modeling in 3D at sixteen, working for clients outside Brazil before writing any code that mattered. That skill never became a hobby — it became Octus Forge, where I now review the work of six modelers across three countries.
Software came later, and came alongside it. Today I build systems that operate on their own: an agent that places orders at a real broker, a detector that finds solar panels in satellite imagery, a pipeline that assembles an entire video with nobody in the middle.
What interests me in all three is not the artificial intelligence — it is what happens when it gets things wrong. An agent with access to money needs a guard that closes by default. A model with suspiciously high metrics needs someone distrustful enough to audit their own test set. A video with numbers on screen cannot let the image generator invent the numbers.
That is why every project here ends in a number: measured by me, with the conditions I measured it under, or marked "from the project" when the measuring was not mine — because a number without conditions is advertising. They live on each project page, where the full accounting fits, not in a headline: what matters is not that the number is big, but that it is true.
How I work
I run a structure of AI agents with separate areas: each front of my work has its own context, a decision log and a backlog, plus a channel where those fronts talk to each other. This site was built inside it.
What I do
Computer vision
I train my own models, measure what they actually get right, and serve them in production — from dataset to inference.
RF-DETRPyTorchONNX RuntimeOpenCV
Satellite-based solar panel detection →CAPTCHA reader for public-data collection →
Agents and automation
MCP servers and skills that take real action, with human approval required before any step that changes state.
Claude CodeMCPFastMCPPython
AI consulting
I go into the operation, find what is repetitive and automate it — including when the right answer is not AI at all, but a macro anyone on the team can maintain.
Claude CodePipedriveOmie ERPTrello
Fourteen agent skills for a finance back office →Reconciling the finance spreadsheet against the ERP →
3D reconstruction and fabrication
From phone to printed piece: photogrammetry, splatting, watertight mesh and resin.
COLMAPGaussian SplattingOpen3DBlender
Software with real money moving
Platforms in production taking real payments, with each franchise’s data isolated inside the database itself.
Next.jsPrismaPostgreSQL RLSAWS
Quantitative research
Hypotheses tested to the end, with costs included — and published even when the answer is that it does not work.
pandasstatsmodelsSciPy