M************h
Profil
Hi! My name is Mohamed, a software engineer with experience contributing to Hugging Face’s core quantization algorithms. I found bugs in their codebase and optimized their CUDA kernels for quantization, achieving a 40% speedup on H100 and B200 GPUs (please see the PR here: https://github.com/bitsandbytes-foundation/bitsandbytes/pull/1746). I worked with Fortune 500 clients at Turing, helped them improve their models with SFT and RLHF datasets, as well as building backend services for reinforcement learning Gyms to train agents on MCP servers and tool calling.
Bildung
Erfahrungen
-Led the development and maintenance of live, containerized Model Context Protocol (MCP) servers for a Reinforcement Learning Gym, replicating public, large-scale APIs to train AI agents on MCP tool calling. -Implemented large-scale public APIs end-to-end (backend, database, MCP), including Jibble, Amazon Seller API, PayPal, and SAP SuccessFactors as standalone Dockerized services. Worked with QA and Prompt Engineering teams to generate agent training datasets using our RL Gym backend replicas for Fortune 500 clients. Implemented verifiers using SQL queries and LLM as a Judge techniques. -Led a team of 4 developers to create supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) reasoning and coding datasets, iterating weekly based on evaluation and research feedback from parallel teams.
Multi-LLM RAG Workflow for Automated Code Documentation Generation LangChain, Weaviate -Developed an agentic Retrieval-Augmented Generation (RAG) Workflow that uses multiple LLMs to extract, synthesize, and generate structured documentation from a FastAPI codebase. Implemented Abstract Syntax Tree (AST) parsing to analyze code structure and automatically generate function-level documentation and dependency summaries. License Plate Recognition on Microsoft HoloLens 2 using Unity 3D Python, C#, Unity 3D, OpenCV, Yolov8 - Built a real-time license plate recognition system for Microsoft HoloLens 2 using Unity 3D, training and deploying a custom YOLOv8 model in Python, integrating live camera feeds via OpenCV, and performing OCR with Tesseract through a Flask API.
Military Soldier
- Built an urban infrastructure simulation in Unity using geospatial and environmental data (OpenStreetMap, satellite imagery, Google Earth, drone scans) to create a digital twin of KAUST City in Saudi Arabia. Modeled and visualized real-world electricity, water, traffic, and resource usage data for infrastructure planning, forecasting, and stress testing. - Collaborated in a 5-person team to deliver interactive Grafana dashboards and automated Plotly reports for real-time urban simulation and analysis (mainly traffic simulations and water/electricity consumption analysis).
Developed VR/AR experiences for HTC Vive, Oculus Quest 2, Windows Mixed Reality, and Microsoft HoloLens 2. Notable projects included a networked VR firearm simulation demo designed for military training purposes on Oculus Quest.
