AI R&D Engineer · MED-EL · Cairo
Ahmed Ashraf Helmi
Research engineer working on biomedical data, computer vision, and on-premises LLM systems — models that stay close to the clinic, the device, and the data they were trained on.
I ship full systems, not just notebooks: real-time vision on constrained hardware, retrieval and SQL generation that never call an external API, and model checks that catch shortcut learning before anything is trusted. Current work at MED-EL sits between research operations and AI on clinical, imaging, and device-generated data.
Currently at MED-EL, Cairo
Selected work
Systems that collect or protect the data, train under real constraints, then check that the model is looking at the right thing.
MED / CV Medical imaging Tumor and COVID classification with explanation checks 96–97% precision GitHub
Problem
High accuracy on a medical model is not enough. The network has to attend to the lesion, not a scanner artifact or the background.
Approach
NASNetMobile for 3-class brain tumor MRI; DenseNet169 for COVID-19 CT. Both audited with Grad-CAM before treating the metric as a result.
Result
96% and 97% precision. Grad-CAM caught a shortcut-learning failure (saliency on background, not the lesion) before the model would have shipped.
LLM / RAG AskDB On-premises natural-language-to-SQL Fully local
Problem
Existing NL-to-SQL tools send schema and query text to a third-party API. That is a non-starter for regulated or private data.
Approach
Hybrid BM25 + cross-encoder re-ranking, a self-correcting SQL loop, and local Qwen2.5-Coder / Llama 3 via Ollama. Nothing leaves the machine.
Result
Six-service Docker Compose stack: FastAPI, PostgreSQL + pgvector, Redis, self-hosted Langfuse, with query tracing and Postman end-to-end checks.
EDGE / CV PlanTech On-device plant disease diagnosis 98% · ~37ms GitHub
Problem
Field diagnosis has to run in real time on low-power hardware. Cloud round-trips are too slow and often unavailable.
Approach
Led a 3-person team. Grew the dataset from 300 to 1,100+ images per class (6,000+ total; 3→5 disease classes) with augmentation that mimics field conditions.
Result
MobileNetV3Large at 98% accuracy, ~37ms inference, deployed with TensorFlow Lite. Graduation project — Grade A+.
DET / OCR ALPDR Arabic license-plate detection and recognition 99% mAP50 GitHub
Problem
Arabic plates need detection and character recognition that hold up on live video, not only still frames.
Approach
Fine-tuned YOLOv11m separately for plate localization and character recognition. Tracked runs in Weights & Biases; EasyOCR as a recognition path.
Result
99% mAP50 plate detection, 98% mAP50 character recognition, shipped as a Streamlit app for images and live video.
Experience
- Clinical: trained in audiology and objective measurements; programmed audio processors; observed about 25 cochlear implant surgeries.
- Research: reviewed papers on hearing preservation and deep-learning segmentation; analyzed study data statistically.
- AI: built semantic segmentation skills with PyTorch, MONAI and Auto3DSeg.
- Taught AI and game development to 6th-grade students: data structures, control flow, and first ML ideas (regression, clustering) without the jargon.
Training
- Built the Arabic license-plate detection and medical-imaging pipelines shown in Selected work.
- Used GAN-based augmentation to expand medical datasets.
- Led a team of 3 on HeartMind AI, a dual wearable (EEG + ECG) health-monitoring concept, including the technical pitch and go-to-market plan.
- Finished 4th of 12 finalist teams at the SIC hackathon.
- Built lane-line detection for autonomous driving with Canny edges and the Hough transform.
- Cleaned and visualized messy tabular datasets in Python.
Skills
Research practice, models, then the systems that serve them.
Research
How I check a model is right
- Dataset curation
- Model validation
- Grad-CAM / XAI
- Experiment tracking
- W&B
- Langfuse
Computer vision
Training and deploying vision models
- PyTorch
- TensorFlow
- Keras
- OpenCV
- YOLO
- TFLite
- EasyOCR
LLMs & retrieval
Private, on-premises pipelines
- LangChain
- Ollama
- pgvector
- BM25 + cross-encoder
- Self-correcting SQL
Systems & data
The plumbing around the models
- Python
- SQL
- FastAPI
- Docker
- PostgreSQL
- Redis
- Pandas
- Scikit-learn
Background
Mansoura University, Faculty of Engineering
B.Sc. in Electronics and Communications Engineering
Sep 2019 — Jul 2024 · Mansoura, Egypt
Courses
- Deep Learning Specialization — DeepLearning.AI / Stanford
- Machine Learning Specialization — DeepLearning.AI / Stanford
- Python for Everybody — University of Michigan
Recognition
- Kaggle Notebooks Expert
- HeartMind AI — SIC Hackathon, 4th of 12 finalist teams
Led a team of 3 on a dual wearable (EEG + ECG) health-monitoring concept, including the technical pitch and a go-to-market plan.
Volunteering
- IEEE MansCSC — Data Science. 3rd, Advanced, Member of the Month; regression notebook R² 0.983.
- Multimedia for IEEE MansSB, Momentum, FEMU, ICPC Mansoura, BreakinPoint.