ahmed@portfolio:~$ whoami
> AI/ML engineer · Egypt → Germany · currently arguing with a loss function
Hey, I'm Ahmed 👋
Software Engineer / AI Master Student
// grounded with my real profile data, powered by OpenAI GPT-4.
About Me
27 · Egypt → Germany
Projects
6 case studies · latest: GPT-Slowrun Ablation
Skills
19 skills, all with receipts
Contact
x4ahmed.omran@icloud.com
Resume
Updated Aug 2026 · PDF
Articles
7 write-ups, 2 platforms
Hypothesis Space
ahmed@portfolio:~$ ls ~/projects▌
Challenging myself into bridging the gap between abstract theory and practical implementation.
State Estimation of a Quadrupedal Robot using Temporal Convolutional Networks
Experimental research investigating the application of Temporal Convolutional Networks (TCN) for accurate state estimation and localization of quadrupedal robots using proprioceptive sensor data, addressing challenges in odometry drift and real-time performance.
Ablation Study of a NanoGPT Slowrun Model: Optimizer, Architecture, and Scaling
A controlled ablation and scale-transfer study of a NanoGPT Slowrun recipe: screening optimizer, hyperparameter, and architecture choices on a single laptop GPU, then testing whether the small-scale winners survive scaling to a 317M-parameter model trained on 8xH100 GPUs.
An Experimental Study of Task-Oriented Dialogue Systems
Six connected experiments spanning the task-oriented dialogue stack: word embeddings, a modular ConvLab-3 pipeline, a fine-tuned BERTNLU model, TripPy dialogue-state tracking, a PPO reinforcement-learning policy, and LLM-based dialogue agents, evaluated on MultiWOZ 2.1.
3D Reconstruction from 2D Brain Scan Slices
A classical computer-vision pipeline that aligns a stack of unregistered 2D brain histology slices into a coherent volume, combining manual and SIFT-based image registration, from-scratch image-filtering implementations, and RANSAC-based robust transformation estimation.
Cell Segmentation and Detection in Fluorescence Microscopy
A comparison of three approaches to nuclei segmentation on the BBBC039 fluorescence microscopy dataset — classical Otsu thresholding, Watershed-based instance segmentation, and a pretrained Contour Proposal Network — evaluated with Jaccard, F1, and Panoptic Quality.
Stereo Scene Reconstruction: Disparity, Depth, and Surface Extraction
A stereo vision pipeline that estimates the fundamental matrix and rectifies an image pair, computes disparity via block matching with left-right cross-checking, converts disparity to metric depth, smooths the result with a Markov Random Field, and extracts isosurfaces with marching squares.