Cairo, Egypt — open to AI / ML / fintech roles
Ahmed Ashraf Helmi
AI Engineer building on-premises, privacy-first ML and LLM systems — from computer vision on medical images to natural-language interfaces over private data.
Ahmed ships full systems, not just notebooks — from real-time computer vision on embedded hardware to on-premises LLM pipelines that never call an external API. Recent focus: retrieval-augmented generation, self-correcting SQL generation, and privacy-first infrastructure for regulated data.
Work
AD-01 AskDB On-premises natural-language-to-SQL engine GitHub ↗
Problem
Teams want to ask databases questions in plain language, but every existing tool ships schema and query data to a third-party API — a non-starter for regulated or sensitive data.
Approach
Built a RAG pipeline with hybrid BM25 + cross-encoder re-ranking and a self-correcting SQL generation loop, serving Qwen2.5-Coder and Llama 3 locally through Ollama — nothing leaves the machine.
Result
A 6-service Docker Compose stack (FastAPI, PostgreSQL + pgvector, Redis, self-hosted Langfuse) with full query tracing, validated end-to-end with Postman.
PT-02 PlanTech AI-enhanced robot for plant health analysis GitHub ↗
Problem
Early plant disease detection needs to run in the field, in real time, on low-power hardware — not in the cloud.
Approach
Led a 3-person team through 2 faculty advisors to grow the dataset from 300 to 1,100+ images per class (6,000+ total, 3→5 disease classes), applying augmentation to mimic real field conditions.
Result
Fine-tuned MobileNetV3Large to 98% accuracy at ~37ms inference, deployed full-stack via TensorFlow Lite for real-time, on-device diagnosis. Graduation project — Grade A+.
AL-03 ALPDR Arabic license plate detection & recognition GitHub ↗
Problem
Arabic plates need detection and character recognition that hold up on live video, not just still images.
Approach
Fine-tuned YOLOv11m separately for plate localization and character recognition, tracked experiments in Weights & Biases.
Result
99% mAP50 for plate detection, 98% mAP50 for character recognition, shipped as a Streamlit app handling both images and live video streams.
MI-04 Medical Imaging AI Brain tumor & COVID-19 classification GitHub ↗
Problem
High accuracy on a medical model isn't enough on its own — it has to be looking at the right thing for the right reason.
Approach
Built NASNetMobile for 3-class brain tumor detection from MRI and DenseNet169 for COVID-19 detection from CT, then audited both with Grad-CAM.
Result
96% and 97% precision respectively — and the Grad-CAM audit caught a shortcut-learning failure mode (saliency on background, not the lesion) before it shipped.
Experience
- Manage investigational sites end-to-end — from study initiation through close-out — while integrating clinical, imaging, and device-generated data for research analyses.
- Develop and automate research data workflows (Excel/VBA, Python) and support technical prototyping of research tools.
- Prepare datasets and contribute to the development and validation of AI/ML and deep learning models on biomedical and device-generated data.
- Designed and delivered an AI and game-development curriculum for 6th-grade students, translating data structures, conditionals, loops, and core ML concepts (regression, clustering) into beginner-friendly lessons.
Education
Mansoura University, Faculty of Engineering
B.Sc. in Electronics and Communications Engineering
Mansoura, Egypt
Training & Internships
Samsung Innovation Campus (SIC)
AI & Data Science Trainee
Jul 2024 — Oct 2024
Built the flight-delay, license-plate, and medical-imaging pipelines that became the Work section — plus GAN-based augmentation for medical datasets.
Information Technology Institute (ITI)
Computer Vision Trainee
Jul 2023 — Sep 2023
Built a lane-line detection system for autonomous vehicles using Canny edge detection and the Hough transform.
National Telecommunications Institute (NTI)
AI Trainee
Aug 2023
Applied data cleaning and visualization techniques in Python to extract insights from complex datasets.
Skills
Courses & Volunteering
Courses
- Deep Learning Specialization — DeepLearning.AI / Stanford
- Machine Learning Specialization — DeepLearning.AI / Stanford
- Python for Everybody — University of Michigan
Volunteering
- IEEE MansCSC — Data Science Team Member. Ranked 3rd (Advanced level, Member of the Month); built a regression notebook scoring R² 0.983.
- Video Production & Graphic Design — Multimedia Creator for IEEE MansSB, Momentum, FEMU, ICPC Mansoura, and BreakinPoint.
Awards
Samsung Innovation Campus Hackathon — HeartMind AI
Led a team of 3 to design a dual wearable (EEG + ECG) health-monitoring system and pitched the technical solution plus a go-to-market and revenue plan to a 5-judge panel.
12 finalist teams
Nov–Dec 2024