Built a custom agent framework inspired by Smolagents to handle complex regulatory queries. Designed a modular orchestrator coordinating tools (retrieval, synthesis, reasoning) through multi-step planning, with Gemini for language understanding and fully extensible agent logic.

Based on Gemini API

Developed a foundation model for spinal MRI reconstruction to support surgical planning. Used a self-supervised 3D Masked Autoencoder architecture (based on Vision Transformers) to learn detailed anatomical features from unlabeled data, which were fine-tuned for spinal root segmentation and artifact correction.

Based on ViT Masked Autoencoders

Created an AI tutor that explains STEM concepts interactively. Fine-tuned a language model using PEFT, SFT, DPO, and RLHF techniques to improve alignment with academic standards and student expectations. Delivered accurate, step-by-step solutions to support autonomous learning.

Based on OpenAI GPT2

Built deep learning models to decode arm movement intentions from ECoG brain signals. Explored architectures like EEGNet and CEBRA to capture motor dynamics, contributing to neurotechnologies for spinal cord injury rehabilitation.

Based on EEGNet & CEBRA

Designed a relaxing factory-building game where players automate furniture production. The game evolves from structured tasks to a creative sandbox, with unlockable buildings and mechanics. Built entirely from scratch, including code, gameplay design, and pixel art.

Based on Shapez.io & Factorio