I build systems that use language models as a component rather than
the whole product: multi-agent pipelines, retrieval over real documents, and the
plumbing that keeps both honest. Most of the work below comes with its architecture
drawn out, because how a system is wired is usually the part worth arguing about.
I'm in my final year of a B.Sc. in AI and machine learning at Sadat
Academy, class of 2027. I'm looking for machine learning engineering internships and
junior roles, and I'm happy to talk about anything on this page.
Eight stages that carry a patient record from intake through medical
NLP, risk prediction, evidence retrieval, drug-safety and guideline checks, and clinical
reasoning to a final report. It runs locally on a fine-tuned Qwen2.5-0.5B with a LoRA adapter.
The risk model reaches 85.2% accuracy and 0.924 ROC-AUC
on UCI Heart Disease (Cleveland, 303 rows). The dataset mirror I started from had
inverted target labels — I caught it by checking feature correlations against what the
domain predicts, then fixed and documented it.
Classifies customer complaints by intent and sentiment, clusters them
by topic, and drafts a suggested reply — all in one Streamlit dashboard a support lead
can actually sit in front of.
Answers questions over a set of specialized documents. They're chunked
and embedded into a vector store, retrieved per question, and passed to an LLM that can
only answer from what came back.
Stack Overflow data in BigQuery, fine-tuned on Vertex AI, orchestrated
with Kubeflow Pipelines, and served through FastAPI with safety and citation checks
standing in front of every answer.
A seven-class dermatoscopic image classifier trained on HAM10000 with a
ResNet18 transfer-learning baseline, Grad-CAM heatmaps so a reviewer can see what the
model looked at, and a FastAPI backend.
Classifies news articles as real or fake on the WELFake dataset using
TF-IDF features and a Random Forest, at roughly 96% accuracy. A FastAPI endpoint
returns the verdict with a confidence score rather than a bare label.
Experience — internships, training, and two hackathon finals
Hindawi Tips
Jun – Jul 2026
LLM & Agents Intern
Built a RAG-based agent and a YouTube video summarization pipeline as part of the
team's LangChain projects, from pipeline design through to deployment.
NTI Summer Training — NLP track
Jun – Jul 2026
Completed
Embeddings, Transformers, LLMs, and fine-tuning. The team capstone became the
customer support assistant above.
IEEE ICIAAI 2026, Damietta
Jul 2026
Finalist — top 37 of 85 teams
Smart Bathroom Safety System. IoT sensing plus a model that flags hazards
such as carbon monoxide, LPG, and low oxygen. I trained the model and built the SHAP
and LIME explanations and the Streamlit dashboard.
EVA AI Hackathon
Feb 2026
Finalist — agent engineer
A RAG compliance agent for EVA Group, built on Python, Pinecone, the Groq API, and
n8n, retrieving over the company's real compliance documents.
Study — degree and certifications
Sadat Academy for Management Sciences
B.Sc. Computer Science, major in Artificial Intelligence & Machine Learning.
If you're building something in AI that needs someone who will read the failure cases
as carefully as the demo — a multi-agent platform, a grounded retrieval system, a pipeline
that has to be defensible — I'd like to hear about it.
