I implement research papers from scratch and build reproducible ML pipelines.
Boules Ashraf Samir, machine learning engineer in Cairo, Egypt. I work on computer vision and NLP with PyTorch.
| Work | Result | Reference point |
|---|---|---|
| CVPR 2016 group activity recognition | 90.5% accuracy | 7% above the reported baseline |
| ECCV 2018 relational network | 89.2% accuracy | Matches the paper, no reference code |
| ECCV 2018 inference time | 50% lower | Vectorized tensor operations |
About
I graduated in Computer Science from Egyptian E-Learning University in July 2026. My work centers on reimplementing published architectures, testing them with ablation studies, and packaging experiments so they can be reproduced. I completed a machine learning internship at FlyRank AI, mentor students at Google Developer Groups (GDG EELU), and work in Arabic and English.
Technical skills
- Languages
- Python, SQL, C++
- Machine learning
- scikit-learn, regression, classification, clustering, anomaly detection, imbalanced classification, feature engineering, model evaluation, ablation studies
- Deep learning
- PyTorch, Torchvision, Transformers, BERT, CLIP, ResNet-50, LSTM, graph neural networks, self-attention, contrastive learning, LoRA/PEFT
- Computer vision
- Video understanding, group activity recognition, object detection (YOLO), image captioning, visual question answering
- NLP
- Masked language modeling, embeddings, intent classification, vision-language models
- MLOps and data
- MLflow, DVC, Docker, FastAPI, GitHub Actions, CI/CD, Git, Pandas, NumPy, BigQuery
Featured projects
Applied ML pipelines. Select a project to open its case study.
Applied ML pipelineNYC taxi trip duration predictionR² of 0.69 on 1.4M+ trip records
- Problem
- Predict how long a taxi trip will take.
- Data
- 1.4M+ trip records.
- Features
- Geospatial and temporal feature engineering.
- Methodology
- Data-leakage prevention and reproducible experimentation across an end-to-end pipeline.
- Result
- R² of 0.69.
- Stack
- Pythonscikit-learnPandasNumPyMLflowDVCDockerFastAPIGitHub Actions CI/CD
- Code
- GitHub repository
Applied ML pipelineCredit card fraud detection0.97 ROC-AUC on 284K+ imbalanced transactions
- Problem
- Detect fraud in a highly imbalanced dataset.
- Data
- 284K+ transactions.
- Models
- A Random Forest classifier, benchmarked against a custom PyTorch MLP.
- Results
- Random Forest: 0.97 ROC-AUC, 0.93 precision, 0.87 F1. The MLP reached 0.80 F1.
- Stack
- PythonPyTorchscikit-learnPandasNumPy
- Code
- GitHub repository
Research implementations
Architectures reimplemented from scratch in PyTorch and checked against published results.
Research implementationHierarchical deep temporal model, CVPR 201690.5% accuracy, 7% above the reported baseline
- Objective
- Reimplement the group activity recognition model from scratch and compare it with the reported baseline.
- Architecture
- Video input
- ResNet-50
- Group-LSTM
- Group activity
- Methodology
- Nine targeted ablation studies on architectural components and training strategies.
- Result
- 90.5% accuracy, 7% above the reported baseline.
- Stack
- PythonPyTorchTorchvisionResNet-50Group-LSTM
- Code
- GitHub repository
Research implementationHierarchical relational network, ECCV 201889.2% accuracy without reference code, 50% faster inference
- Objective
- Reproduce the paper’s spatial-temporal relational architecture from scratch, without reference code.
- Architecture
- ResNet-50
- Graph neural network with message passing
- Group activity
- Optimization
- Vectorized tensor operations for feature aggregation reduced inference time by 50%.
- Result
- 89.2% accuracy, matching the paper’s benchmark.
- Stack
- PyTorchResNet-50Graph neural networksRelational modeling
- Code
- GitHub repository
Research implementationFrom-scratch Transformers110M-parameter BERT, CLIP, and a long-context Transformer
- Objective
- Build foundation-model components from first principles in PyTorch.
- Implementations
- nanoBERT: a 110M-parameter BERT pretrained from scratch with masked language modeling.
- CLIP: ViT image encoder, causal Transformer text encoder, symmetric InfoNCE loss, and a learned temperature.
- Long-context Transformer: grouped-query attention, YaRN RoPE, FlashAttention-style chunked attention, sliding-window and global layer interleaving, attention sinks, and KV caching.
- Validation
- Correctness tests verify numerical equivalence with PyTorch SDPA.
Professional experience
Machine learning intern, FlyRank AI
Jul 2026 – Sep 2026
- Developed embedding-based clustering and intent-classification workflows on real-world search and analytics data to support search understanding.
- Prepared large-scale analytics datasets for modeling by handling missing fields, joining sources, and flattening nested structures with Python, Pandas, SQL, and BigQuery.
AI instructor and mentor, Google Developer Groups (GDG EELU)
- Delivered ML and deep learning workshops to 30+ students and led Python code reviews and project debugging.
Programming instructor, Code Kids Academy
- Taught programming fundamentals to 48+ students through practical exercises and self-developed projects.
Participant and mentor, Egyptian Collegiate Programming Contest (ECPC)
Certifications
- Machine Learning Diploma
- Artificial Intelligence
- AI & Data Science: Generative AI Professional
- Deep Learning Diploma (in progress)
- Data Structures and Algorithms (in progress)
Contact
Open to junior and applied machine learning engineer roles in Egypt and remote.