Building the infrastructure layer for multi-agent AI systems.
I design systems where AI agents built on different frameworks — LangChain, AutoGen, CrewAI, Google ADK — discover and collaborate with each other at runtime, without point-to-point adapters. My current focus is agent interoperability, distributed orchestration, and making small models do what people think requires 90B parameters.
Currently finishing my Computer Engineering degree at PUC-Campinas and working as an AI Engineer at IBM. Fueled by coffee and supervised by cats. ☕🐈
Axon — open-source infrastructure for multi-agent AI. Agents register once. Any workflow can use them. Dynamic discovery via A2A + MCP, with UCB1-based gateway routing that learns which resources work best over time.
pip install axon-framework
axon init && axon ga serve
axon pa run --query "your task here"Based on my Bachelor's thesis — 7 experiments, validated end-to-end across Gateway and Principal Agent components.
- 👩💻 AI Engineer @ IBM Research (2024–2025) and IBM Consulting (2025–present)
- 🚀 Deployed HuggingFace LLMs with vLLM on Red Hat OpenShift — the setup later benchmarked IBM vs NVIDIA hardware for LLM workloads
- ⚡ Cut IVR analysis time from 1–2 business days to under 3 hours with a Python/Ollama pipeline — became a reusable asset across IBM LATAM consulting teams
- 🏆 Made it to the Cubo Itaú finals in B3's entrepreneurship competition — out of 500+ applicants
- 💪 Helped revive the competitive programming team at PUC-Campinas (and convinced my professor to fund us 😄)
Before I was building agent frameworks, I was building in other ways.
I was one of the first people to publish tutorials on how to create custom skins in KoGaMa — a Minecraft-style browser game — back when the community was still figuring it out. Thousands of kids used those tutorials. That was my first experience teaching technical concepts to strangers on the internet, and I haven't stopped since. 🎮
I also spent years doing 3D modeling in Blender 🎨 — it taught me something that translates directly to systems design: the difference between something that looks correct and something that is correct under the surface.
I write about the internals of AI systems — not tutorials, but the design decisions and tradeoffs that don't make it into documentation.
- 🎯 Why I used UCB1 to route between AI agents (coming soon)
- 🔬 Implementing multi-head attention from scratch (Kaggle notebook)
The tools I reach for, roughly in order of frequency:
Python PyTorch FastAPI LangGraph A2A MCP Ollama vLLM Red Hat OpenShift Docker
Frameworks I've built things with: AutoGen LangChain CrewAI HuggingFace
"Sometimes it is the people no one imagines anything of who do the things that no one can imagine." — Alan Turing 🖤