At Unicamp's HIAAC, I design a multi-agent architecture for LLMs that combines dynamic memory retrieval with inter-agent communication. I benchmark approaches including ReAct, chain-of-thought, and task decomposition to evaluate their effect on task success.
At BOVIFOCR, I built the submission that placed first in Sub-Task 2.1 of the FRCSyn Challenge at WACV 2024. I also trained and evaluated face recognition models on synthetic identity datasets and measured their performance across eight benchmarks.
My research on synthetic facial data led to a SIBGRAPI 2024 publication and a survey submitted to IEEE Access. I also authored a masked face recognition benchmark presented at SIBGRAPI 2022.
