# Mohamed Thabet

> Machine learning and simulation engineer. I'm 19 and I learn by rebuilding things from first principles: neural networks, decision trees and clustering in raw NumPy, and gravity you can throw planets into. If I can't write it from scratch, I don't consider it understood yet.

I started with physics simulations: a ball thrown with drag and bounce, a solar system in VPython, and a Pygame window where you slingshot planets around a sun. That turned into machine learning, where I write every algorithm by hand before touching a library. My MNIST network is NumPy matrix maths with hand-derived gradients, and switching ReLU to LeakyReLU took it from 10% to about 80% accuracy. I also do exploratory data analysis on real datasets, such as 187,000 house listings, using Pandas and Seaborn. At Samsung Innovation Campus I studied AI and machine learning, and in a team of six built Bayan, an Android app that simplifies Arabic text for readers with dyslexia.

## Links
- Website: https://mohamedthabet.com
- All projects: https://mohamedthabet.com/projects
- Machine-readable profile (JSON): https://mohamedthabet.com/api/profile
- Github: https://github.com/mohamedthabet7
- Linkedin: https://www.linkedin.com/in/mohamed-thabet-269a44441/
- Email: abdwadood2000@gmail.com

## Focus areas
- Machine learning from scratch
- Neural networks
- Backpropagation
- Decision trees
- Clustering
- Regression and statistics
- Exploratory data analysis
- Physics simulation
- NumPy
- Python

## Projects
### Bayan (بيان)
Samsung Innovation Campus capstone with a team of six: an Android app that simplifies hard Arabic text, adds diacritics and reads it aloud, all on the phone. I worked on the app.
- Built the PROCESS_TEXT action, so «تبسيط» appears in the text-selection menu of any app, and its result sheet
- Rebuilt the app in Jetpack Compose and Material 3 Expressive, with model download, settings and result states
- Simplify → diacritize → read aloud with the spoken word highlighted, fully offline with ONNX Runtime
- Data annotation and quality review; the best model keeps the meaning on 82% of the test set
Tech: Kotlin, Jetpack Compose, ONNX Runtime, Python
Live: https://bayan-android.vercel.app
GitHub: https://github.com/Cogni-Libya/Bayan

### Planets Through Space
A Pygame gravity sandbox: drag to aim, release, and Newton's law decides whether the planet orbits, escapes or falls in.
- Drag-and-release slingshot launch, with the speed set by the drag length
- Newtonian gravity F = G·m₁·m₂ / r² from a sun of 1.989 × 10³⁰ kg
- Euler integration of velocity and position at 400 steps per second
- Planets are removed when they leave the screen or hit the sun
Tech: Python, Pygame
Live: https://mohamedthabet.com/#gravity
GitHub: https://github.com/mohamedthabet7/Practice-Projects/blob/main/PlanetThroughSpace-Project

### MNIST Net
A two-layer network that reads handwritten digits, written in raw NumPy with hand-derived backpropagation.
- 784 → 10 → 10 architecture with a numerically stable softmax
- Analytical gradients for every weight and bias, no autodiff
- Diagnosed the dying-ReLU problem stuck at 10% accuracy
- LeakyReLU (α = 0.1) lifted test accuracy to about 80%
Tech: Python, NumPy, Jupyter
GitHub: https://github.com/mohamedthabet7/Machine-Learning-Algorithims

### ID3 Decision Tree
Shannon entropy and information gain worked by hand, then automated into a recursive ID3 builder that prints the tree.
- Manual entropy and information-gain trace for every branch
- Recursive ID3 that picks the split with the most information gain
- Detects pure leaves where H(S) = 0
- Renders the learned Play Tennis tree as clean ASCII
Tech: Python, Pandas
GitHub: https://github.com/mohamedthabet7/Machine-Learning-Algorithims

### K-Means vs K-Medoids
Both clustering algorithms from first principles, compared on how they handle noise and outliers.
- Euclidean distance and centroid updates written out in NumPy
- PAM medoid search: the centre must be a real data point
- Tested on an 8-point set and a 40-point synthetic set
- Shows why medoids resist outliers better than means
Tech: Python, NumPy
GitHub: https://github.com/mohamedthabet7/Machine-Learning-Algorithims

### California Housing
OLS, SGD, Lasso and Ridge compared on California house values, with a VIF diagnosis of correlated features.
- Correlation heatmaps and distributions with Seaborn
- Four regressors compared: OLS, SGD, Lasso and Ridge
- Found AveRooms and AveBedrms correlated at about 0.85
- VIF = 1 / (1 − R²) used to drop redundant predictors
Tech: Python, scikit-learn, Pandas, Jupyter
GitHub: https://github.com/mohamedthabet7/Machine-Learning-Algorithims

### House Prices EDA
Exploratory analysis of a 20-column housing dataset: cleaning, missing data and a price column with a skew of 177.
- Duplicate and missing-value audit across all 20 columns
- Renamed and retyped mislabelled object columns
- Measured extreme right skew and planned a log transform
- Binned prices into intervals to size the outliers
Tech: Python, Pandas, Seaborn, Jupyter
GitHub: https://github.com/mohamedthabet7/Data-Analysis-and-EDA

### Solar Simulation
Planets that pull on each other in a VPython 3D scene, with collisions and labelled bodies.
- Mutual gravity between every pair of bodies
- Launch angle and speed set per planet in 3D
- Collision detection when bodies touch
- Live 3D canvas with a reference grid at 60 FPS
Tech: Python, VPython
GitHub: https://github.com/mohamedthabet7/Practice-Projects/blob/main/SolarSimulation.py

### Ball Throw
A 3D ball thrown at any angle, slowed by drag and friction, bouncing with set elasticity.
- Launch from vertical and horizontal angles in 3D
- Gravity of 9.8 m/s² applied each frame
- Drag, friction and elasticity coefficients
- Comes to rest on the floor instead of jittering
Tech: Python, VPython
GitHub: https://github.com/mohamedthabet7/Practice-Projects/blob/main/BallThrow.py

### Perceptron
A single neuron that learns logic gates, starting from random weights and stopping at zero error.
- Forward perceptron with a bipolar step function
- Self-learning loop with Rosenblatt's update rule
- Random initial weights between −5 and 5
- Trains until total error on the truth table is zero
Tech: Python
GitHub: https://github.com/mohamedthabet7/Machine-Learning-Algorithims

## Certifications
- Artificial Intelligence course, capstone: Bayan, Samsung Innovation Campus