Projects

My work spans robot learning, embodied AI, generative models, and machine-learning systems. This page highlights the projects most closely related to my current research, followed by selected earlier software work.

1. Countdown GRPO: I trained Qwen3.5-0.8B-Base using reinforcement learning (GRPO), without any supervised fine-tuning, to solve Countdown arithmetic puzzles on a single 16 GB laptop GPU. In one 300-step LoRA run, I improved held-out accuracy from 0% to 33%. I also ran a controlled three-seed experiment to study reward design and found that adding a seemingly “helpful” partial-credit reward actually reduced average accuracy from 23% to 6%.

• GitHub: https://github.com/mohamedsobhi777/countdown-grpo
• W&B: https://wandb.ai/mohamedsobhi777/countdown-grpo
• Report: https://github.com/mohamedsobhi777/countdown-grpo/blob/main/report/main.pdf

**2. Morpha - AI App Builder
Morpha started as an AI app builder that turned founder intent into research, PRDs, implementation plans, generated code, and a working product surface. The thesis was that AI should compress the path from an idea to execution.

**3. OpenMarble - Open Source alternative to Marble from World Labs.

MarbleOS is an experiment in what a personal spatial OS might look like if it were built around generative 3D. Upload a photo, and MarbleOS reconstructs it as a navigable Gaussian Splat - viewable, shareable, and editable right in the browser. The interface takes its aesthetic cues from visionOS: glassmorphism, depth, spring animations, and a window-based app model.

Github: https://github.com/mohamedsobhi777/OpenMarble

Check out my github for more: https://github.com/mohamedsobhi777