MSA University × University of Greenwich · 6th of October, Egypt · Graduating June 2027
I build LLM-based systems for software engineering and security, organized around a single question: how do you make a model's output reliable enough to trust in constrained, safety-critical settings? My work runs from formal evaluation of LLM-generated code, to agentic pipelines for industrial threat intelligence, to efficient inference on limited hardware.
In summer 2026 I am joining the University of Illinois Summer Research in Software Engineering (UIUC++ SRSE) 2026 cohort to work on a code generation research project, after being selected through a Verus-based formal-verification evaluation. Earlier in 2026 I was an AI Intern at ControlPoint, an OT/ICS cybersecurity startup, where I designed agentic architectures for industrial threat intelligence. At MSA University I lead the research group at the CSE Outstanding Hub and teach C++ to secondary school students.
- LLMs for code generation. Reliable and secure code synthesis for non-trivial tasks, evaluated formally and empirically rather than by demo.
- Agentic AI systems. Architecture, reliability, and evaluation of autonomous LLM pipelines: task decomposition, tool use, and output contracts.
- Efficient inference. Quantization and principled quality-versus-cost tradeoffs for deployment on constrained hardware.
- AI for cybersecurity. Vulnerability analysis and automated threat intelligence for OT/ICS environments.
| Role | Where | Focus | |
|---|---|---|---|
| 🔬 | Incoming Undergraduate Researcher | UIUC++ SRSE 2026 | Code generation research; selected through a Verus-based formal-verification evaluation |
| 🧠 | Head of Research | Outstanding Hub, CSE Dept., MSA University (since Nov 2025) | Leading a research group and survey on automation in AI/ML systems |
| 🖥️ | Programming Instructor | 6th of October, Egypt (since Oct 2025) | C++ fundamentals, secondary school level |
🔒 Automated OT/ICS Threat Intelligence Pipeline
An agentic pipeline for industrial cybersecurity. It polls the National Vulnerability Database for live CVE feeds, applies a two-stage classifier (a domain-specific keyword prefilter, then a single LLM pass under strict JSON output contracts), and surfaces findings through an analyst-facing Streamlit dashboard, deployed via GitHub Actions. This work preceded and shaped my internship at ControlPoint.
🤖 Ouro: Closed-Loop Active Learning System (Team Lead)
A continual-learning system built on Llama 3.2 3B. A human-in-the-loop pipeline captures correction signals at inference time, triggers LoRA adapter retraining, and hot-swaps weights with no service interruption, so the model improves in production without redeployment. Trained on a 13k-row curated dataset (DailyDialog plus a mental-health corpus), it reaches 4.08 perplexity and 0.88 BERTScore F1 on a held-out set. The choice of 3B over 8B came from a controlled comparison under an identical protocol, documented as an efficiency-versus-quality tradeoff under 4-bit quantization for edge deployment. MLflow tracking and GitHub Actions throughout.
Languages
Machine Learning and LLMs
MLOps and Systems
A Mathematical Breakdown of Gradient Descent · CSE Innovation Hub, MSA University · Feb 2025
A derivation of gradient descent from first principles: partial derivatives of the loss function, the effect of learning rate on convergence geometry, and the Batch, Stochastic, and Mini-Batch variants.
Open to research collaboration in LLM systems, agentic AI, and software engineering for AI.