I build systems that learn to act, not just predict. My focus is objective-driven learning — training models through execution feedback instead of imitation, multi-agent pipelines that reason through explicit intermediate state, and causally-aware memory that scales linearly without quadratic attention blowup. I ship end-to-end: reward pipeline design, inference-time planning, constrained report generation, and production deployment in resource-constrained clinical environments.
Co-founding MIOSA and leading AI at Veredictos. Also building PaymentsLine. Senior AI/ML Engineer at Lunivate. Started consulting at 15.
| Achievement | Details |
|---|---|
| 1st Place — CrewAI Agentic AI Challenge 2025 | Built DiabetesEye with Gabriel Maia. Only Brazilian team in the global top 10. Judges cited the combination of agentic AI, classical ML and low-level API control as the differentiator. |
| AWS Machine Learning Specialty Certified | Amazon Web Services, 2024 |
| Finalist — Brazilian Nuclear Olympiad 2025 | IRRAD[IA]: YOLOv8-based AI system for food irradiation control with Cobalt-60 dose optimization and industrial safety loop |
| Arc × Circle Hackathon 2026 | BaaS Arc: autonomous 4-division banking syndicate, 110 production functions, DeFi, sub-second RTP/FedNow settlement |
| MIOSA — 7k downloads in 3 days | Word of mouth only. Built on OSA, a local-first Elixir/OTP agent runtime |
Veredictos — Retinal AI platform screening for diabetic retinopathy, glaucoma, and hypertensive retinopathy from fundus images. The model (86.8M params, ViT + MedGemma) passed GATE 1 with 94%+ sensitivity across all three diseases, trained on 24K images from 18 datasets. Active pilots with Prefeitura do Rio de Janeiro (SUS) and Centro Carioca do Olho. Backed by NVIDIA Inception and Google for Startups.
PaymentsLine — Payments infrastructure for AI-native businesses, supporting clinical billing and agentic transaction flows.
MIOSA — Self-learning agent platform built on OSA, a production Elixir/OTP orchestrator with hot-reload skills, PubSub event routing, and Rust NIFs. 7k downloads in 3 days.
| Project | Stack | What it does |
|---|---|---|
| CodeForge ASM | PyTorch · TRL · GRPO | Qwen3.5-2B trained from zero supervised labels to write x86-64 NASM assembly using pure execution signal. Two-phase GRPO on Kaggle T4x2. MCTS over instruction sequences at inference time. |
| hackaton-claw | Python · Claude | Dual-loop autonomous scientific analysis: strategic loop decomposes goals, tactical loop implements each step with retry. All agents run on Claude. |
| MCU | Python | Linear-time O(N) causal memory with built-in causal coherence. Proved epsilon-causal consistency with M = O(D log N) via instrumental variable-based causal distance estimation. |
| AEG | Elixir · OTP | Production agent runtime: guard → plan → memory → model → eval → persistence. Agents plan against a learnable world model before executing. Q-learning RL layer. Adapters for OpenAI, Anthropic, Gemini, Ollama. |
| BaaS Arc | Python · Gemini | Autonomous 4-division banking syndicate (Front Office, Risk, Treasury, Clearing). Every transaction cleared through all four agents. Gemini-powered fraud detection, OFAC/PEP screening, DeFi yield farming, sub-second RTP/FedNow. |
| ASP | Python | LLM prompt injection detection using 3D Gaussian probability fields and morphological intent detection. Minimal dependencies by design. |
RL & Training — GRPO · PPO · MCTS · TRL · LoRA/QLoRA · reward modeling · DQN
Vision & Medical — YOLOv8 · OpenCV · TransUNet · EfficientNet-B3 · UNet++ · MedGemma
Agents — CrewAI · LangChain · LlamaIndex · Elixir/OTP · MoE · RAG · world-model loops
Infra — AWS · Docker · Kubernetes · FastAPI · PostgreSQL · Redis · Vector DBs · MLOps
