Senior Software Engineer focused on AI, shipping AI systems to production. I started in robotics, moved into computer vision for agriculture and medical imaging, and today build LLM agents and retrieval systems. In every case the job has been the same: going from real, unstructured data to a feature running in production.

MSc in Computer Science (UFU) · Mechatronics Engineer · Uberlândia, Brazil · open to remote
Email · LinkedIn · ORCID

LLM systems for legal tech · _{2024 – present}

• Multi-agent assistant in production for 17k+ legal professionals, planning before it acts and asking before it assumes
• Citation provenance: every legal reference must resolve to a source actually retrieved in that turn
• Case-law search over 83M court decisions, with LLM query refinement and a reranking layer
• LLM-as-Judge evaluation and observability on Langfuse and BigQuery, with caps against runaway agent loops

Medical imaging · _{2023 – 2024}

• Lesion segmentation in mammography, from 75% to 85% Dice
• Pediatric bone age estimation, cutting mean absolute error by about 6 months
• Chest X-ray detection with YOLO, DICOM ingestion and queue-based inference pipelines

Precision agriculture · _{2020 – 2022}

• Weed segmentation and classification on drone orthomosaics of 2 GB+
• Spray maps covering only the ~30% of a field that actually needs herbicide

Humanoid robotics · _{2018 – 2021}

• Led motion for a 22-DOF soccer robot competing in LARC and RoboCup
• Migrated the team's ROS stack from Kinetic to Noetic and bridged it to ROS 2

Applied computer vision research at the Federal University of Uberlândia, published in international journals and conferences.

• Detecting nutrient deficiencies in coffee leaves over time, comparing CNNs, Vision Transformers and hybrid models across 98 experiments
• Synthetic image datasets for inspecting insulators on high-voltage power lines
• Multichannel analysis for weed detection in drone imagery

LLMs & agents
Search & data
Vision & ML
Backend & infra
Frontend
Robotics

• trueodds · Small decision model that returns a probability for each option of a question in one forward pass, and measures whether those probabilities are true: calibrated on data like its training data (ECE 0.007), overconfident on tasks it never saw.
  _{PyTorch · ModernBERT · calibration · MLflow}
• DocVault · Document RAG that answers with exact citations linked to the source page, with agentic multi-step retrieval and a fully offline mode.
  _{FastAPI · React · Qdrant · PostgreSQL · Ollama}
• coffee-temporal-selection · Research code showing that an accuracy peak found by selecting images over time belongs to the validation protocol, not the plants: 78% with an image-level split, 37% once plants are kept apart.
  _{PyTorch · ResNet50 · grouped cross-validation}
• coffee-soil-nutrients-dissertation · MSc dissertation code. CNNs, Vision Transformers and hybrid models detecting nutrient deficiencies in coffee leaves over time, explained with Grad-CAM and attention maps.
  _{PyTorch · Vision Transformers · Grad-CAM}
• auto-insulabelling · B.Eng. thesis. Segments insulator strings on high-voltage towers and classifies the contamination on the discs, trained only on synthetic Unity 3D images: 0.92 Dice on real photos.
  _{TensorFlow · U-Net / LinkNet · synthetic data}