BSc in Computer Science, Scientific Computing Department | GPA: 3.3
Experience
ZinadJuly 2024 – Present
AI Engineer (Part-Time)
Built a phishing simulation engine that generates themed emails and working fake landing pages from the parameters a user picks, driving Ollama from a Laravel backend and rendering the emails with MJML.
Wrote an email analyzer in Python, LangChain and Ollama that flags phishing from tone, context and grammatical anomalies; its verdicts feed the platform's security metrics.
Developed an internal RAG system that answers employee questions from the company's policy PDFs, splitting the documents with recursive character chunking with overlap, embedding the chunks into a FAISS vector store and retrieving the passages relevant to each question.
Shipped a chatbot into the live application with LangChain and Ollama, streaming replies token by token instead of making the user wait for the full answer.
Projects
RAG
Built a document Q&A service in FastAPI and LangServe, splitting a company policy corpus into overlapping chunks with a recursive character splitter, embedding them with word2vec, indexing them in a Chroma vector store, and grounding Gemini on the chunks retrieved for each question.
Added a phishing screener that runs an email body through a self-hosted Llama on Ollama, returning nine checks (urgency language, requests for confidential information, character-encoding tricks such as Cyrillic lookalikes, and others) as structured JSON with English and Arabic justifications, covered by end-to-end API tests.
Extended retrieval to databases with a text-to-SQL chain that feeds the live MySQL schema to the model, executes the generated query and phrases the result in plain English.
Classified emotion from raw EEG brain signals on the SEED-IV dataset, tuning nine deep-learning architectures (DaViT, DGCNN, CCNN, BFENet, TSCeption and others) in PyTorch under subject-dependent and cross-subject protocols, reaching 96% accuracy on binary and 85% on 4-class emotion recognition.
Combined the dataset's precomputed differential entropy features with our own: per-band differential entropy extracted through db4 wavelet decomposition, and hemispheric asymmetry (DASM) features derived from it.
Built the raw-signal preprocessing pipeline in SciPy and NumPy (Butterworth bandpass, downsampling, z-score normalization, windowing), and re-ran the experiments on the 14 channels of a consumer Emotiv EPOC+ headset to test viability on affordable hardware.
Blood Cell Detection with YOLO
Located and classified red cells, white cells and platelets in blood-smear microscope images, fine-tuning pretrained YOLOv8 and YOLOv5 detectors on 364 annotated slides and benchmarking the two generations on the same split at 0.93 mAP@50.
Wrote the dataset pipeline behind the training run, converting ~5000 corner-coordinate bounding boxes into normalized YOLO center format with OpenCV.
Evaluated per class, holding 0.98 mAP@50 on white cells despite being outnumbered ten to one by red cells.
Graph-Based Image Segmentation
Split images into regions in C#, implementing a graph segmentation algorithm from its paper over a graph built from every pixel.
Sorted those edges in linear time with a 256-bucket counting sort rather than an O(E log E) comparison sort.
Segmented the red, green and blue channels independently and intersected the three results.
MEC Model of Leadership: Completed an extensive leadership program (160 hours) covering
emotional intelligence, effective communication, self-awareness (50 hours), the Enneagram, problem-solving,
and team building.
