Hi! I'm Haroon — BS Artificial Intelligence, SZABIST Islamabad
I build AI systems end to end: retrieval, computer vision, and deployment. And I evaluate them before I call them done.
BS AI
— ABOUT ME
I'm a BS Artificial Intelligence student at SZABIST Islamabad (CGPA 3.55), graduating in 2027. Most of my work sits in retrieval and LLM systems and computer vision. SupportGenie is a RAG support bot that abstains when retrieval is weak and holds up against jailbreaks; PixSearch is a CLIP-and-FAISS image search engine on Hugging Face Spaces. Alongside my degree I run the production website and SEO for a Dubai business, so I know what keeping a live system up actually involves. I'm looking for an AI/ML internship, and building toward an MS in Germany for Fall 2027.
Battery Boost Dubai
CodeAlpha
A white-label RAG support platform: businesses plug in their own docs and get an assistant that cites its sources. Retrieval below a calibrated confidence threshold never reaches the LLM, so it hands off to a human instead of inventing a refund policy. Hardened against jailbreak and role-swap prompts. 83% Hit@3 on a held-out eval set.
Content-based image retrieval over a 1,000-image corpus. ResNet-50 and CLIP ViT-B/32 embeddings benchmarked against a classical HSV + HOG + ORB baseline, indexed in FAISS for sub-millisecond top-K search on CPU. mAP 0.93.
Plant disease detection from leaf images. A CNN built in TensorFlow/Keras classifies leaves as healthy or one of several diseases, trained with resizing, normalization, and augmentation. 79% test accuracy, with per-class precision and recall reported alongside it.
A native Android budgeting app built with Kotlin, Jetpack Compose, and Room, with home-screen widgets. A side project in shipping outside my ML stack, open-sourced on GitHub.
RAG pipelines that cite their sources, refuse what they don't know, and get measured on a real eval set.
Image retrieval, classification, segmentation. Transfer learning on custom datasets, benchmarked against classical baselines.
FastAPI, Docker, Streamlit, Hugging Face Spaces. Models turned into services and demos people can use.
