Researcher · Fellow Estudar '25 · Computer Engineer @ UFPB
Designing intelligent systems that learn, reason, and optimize.

Artificial Intelligence · Reinforcement Learning · Combinatorial Optimization · LLM-based Reasoning & Agentic Architectures · Multi-Agent Systems · Computer Vision

Undergraduate Thesis — Constructive RL for Parallel Machine Scheduling with Rejection (2025–2026) Advisor: Prof. Anand Subramanian, UFPB

• Constructive PPO agent with a GAT-based GNN encoder + cross-attention decoder for sequential job scheduling on parallel machines with maintenance and rejection constraints.
• POMO-style multi-start rollouts with Thresholded Lexicographic Policy Gradient for strict multi-objective optimization (no scalarization), combined with Self-Imitation Learning.
• Invalid action masking for deadline feasibility + WSPT-based post-processing for near-optimal weighted completion times.

Research Collaboration — Logistics & Optimization Group (LOG), UFPB (Nov 2024 – Jul 2025) Supervisors: Prof. Bruno Fânzeres (PUC-Rio), Prof. Anand Subramanian (UFPB) · Collaborator: Carlos Neves

• MINLP formulation for exact neural network training, with bilinear terms linearized via McCormick envelopes and piecewise-linear activation approximations.
• Developed and benchmarked an exact MILP training procedure for certifiably optimal, interpretable small-scale networks.

Research Fellow — Fiocruz Innovation Incubator (Dec 2023 – Mar 2024)

• Time-series forecasting model for ICU bed demand in Brazil's Federal District; top 10 team nationwide, grant awarded by Fiocruz.

• Kaiky Maia et al. — Integrating Computer Vision into CACTO: A System for Fraud Prevention and Secure Transport Monitoring. WFA @ WebMedia 2025. [paper]
  • Vehicle re-identification system (ResNet backbone + re-ranking) deployed in production for SEFAZ-PB, achieving 12× inference speedup (93.5% Rank-1 accuracy on VeRI), the largest such deployment in Northeast Brazil.

• Former President, TAIL (UFPB) — Feb 2022 – Jul 2026. Led 70+ students across CV, NLP, and RL projects; secured partnership with Canastra Ventures.
• Fellow, Fundação Estudar — Aug 2025. Top 0.25% of applicants nationally.

• AutoOpt — LLM-driven automated heuristic design; closed-loop framework where a local LLM iteratively proposes and evaluates code-level modifications to heuristic solvers, evaluated on CVRP benchmarks.
• ZoomAI — Compositional infinite-zoom image synthesis chaining GPT-3.5, BLIP, and Stable Diffusion 2 Inpainting.
• UG0 — Vision-guided educational robotics: Arduino arm with real-time pencil-tracking via C++/OpenCV.