AI & Machine Learning Solutions

I build machine learning and deep learning models for forecasting, classification and computer vision problems with Python, TensorFlow and scikit-learn. I look at what the data says first, turn the problem into a measurable goal, and start from the simplest model that works.

Two of these studies have been published: the CGU model, which combines a CNN with a GRU, appeared in the Egyptian Informatics Journal (Elsevier), and the dam water level forecasting study was presented at ICHORA 2024 and is indexed in IEEE Xplore. In computer vision I have worked on real-time object detection and tracking with Faster R-CNN and on histopathological image classification with VGG16 and VGG19.

What is included

• Defining the problem and what success means, together
• Deep learning architecture design (e.g. CNN + GRU; model search with NAS and Bayesian optimization)
• Transfer learning from pre-trained models when data is scarce (e.g. VGG16/VGG19)
• Model selection, training and hyperparameter optimization
• Evaluation with metrics that fit the job (e.g. F1, ROC-AUC, R²)
• A clear report of the results and of the model’s limits

Process

1. We define the problem We agree on what to forecast or classify, and on how a result will be judged.
2. I study the data Exploratory analysis shows the data’s quality, its gaps and the variables that drive the outcome.
3. I build and compare models I start from simple baselines and compare stronger methods on the same metrics.
4. I report the result I share plainly how well the model works and where it can go wrong.

Technologies

Python, TensorFlow, Keras, PyTorch, scikit-learn, XGBoost, Pandas, OpenCV

Example projects

• Dam Water Level Prediction with Time Series Models
• Customer Churn Analysis
• CGU (CNN-GRU Unified) Model with NAS-Based Bayesian Optimization
• Real-Time Object Detection & Tracking (Faster R-CNN + RNN)

Frequently asked questions

How much data do I need?

It depends on the problem. Once I have seen your data, I tell you honestly whether it is enough. With little data, transfer learning from a pre-trained model can be a good start; in the histopathology project I used VGG16 and VGG19 pre-trained on ImageNet this way.

Is accuracy enough on its own?

Usually not. On imbalanced data a high accuracy can mislead, so I evaluate with metrics that fit the job, such as F1, precision/recall or ROC-AUC.

How are the timeline and the price decided?

By the scope of the work. A short note on what you need is enough to start; once I have seen the data and we have agreed on what success means, I send you a quote with the timeline and the price.