Yusuf Shihata

ML Researcher | ML Engineer | Mobile Developer

I design and build machine learning systems from scratch. Published my first research paper at 19 — introducing Gated Recursive Fusion, a novel multimodal architecture. Currently building Neura, a custom deep learning framework in CUDA.

About Me

My focus lies at the intersection of systems and intelligence: CUDA-based frameworks, efficient transformers, and novel multimodal reasoning strategies. I aim to make AI more interpretable, efficient, and structurally grounded—not just bigger.

Languages

  • Python
  • C++
  • SQL

AI/ML Libraries

  • PyTorch
  • Hugging Face
  • NumPy
  • Pandas
  • Matplotlib

Tools & Concepts

  • CUDA
  • Docker
  • Git & GitHub
  • System Arch
  • PostgreSQL
  • Linux/CLI

Education

Bachelor's in Artificial Intelligence
Kafr El-Sheikh University, Egypt (Expected Graduation: Jan 2029)
Current GPA: 3.73

Experience

AI Researcher (Independent)

May 2024 – Present

Published Gated Recursive Fusion (GRF), a transformer-based architecture for efficient multimodal fusion. Proposed a recursive and gated approach to reduce memory usage and improve interpretability. Submitted to ACM Multimedia 2025.

View Paper

Open-Source Contributor

Hugging Face — Jun 2024 – Present

Working on expanding core functionality in Transformers: adding features to the core library pipelines, fixing bugs in QA pipelines, refactoring modules, and improving test coverage.

View Contributions

Projects

Gated Recursive Fusion (GRF)

A multimodal transformer that fuses modalities recursively with a gated mechanism. Designed for low-compute environments like robotics and perception stacks.

Tech: Python, PyTorch

GitHub

Neura: DL Framework

A PyTorch-inspired deep learning framework built with Python and CUDA. Includes tensor operations, GPU kernels, and a basic autograd engine for research.

Tech: Python, CUDA, C++

GitHub

PoktAid

Offline-first mobile app that combines speech recognition, computer vision, and LLMs to assist users in medical first-aid situations on-device.

Tech: React Native, TFLite

GitHub

AuditGPT

A RAG system for financial/legal auditing. Integrates document chunking, vector search, and a fine-tuned LLM to answer complex queries from uploaded files.

Tech: LangChain, FAISS

GitHub

Get In Touch

If you're building something serious in AI, need a collaborator, or just want to connect — reach out. I'm always open to new ideas and challenges.