Ayman Abo Ghonim

My GitHub Profile - Software Engineer & Data Science Enthusiast

Single-node local LLM inference platform — KServe + vLLM (IBM Granite 4.2-3B AWQ INT4) on WSL2/k3s with an OpenAI-compatible gateway contract, Gemini tier-2, telemetry, hermetic CI, Argo GitOps.

Minimal CPU-only script that checks whether greedy decoding of a small language model (SmolLM2-135M, Q4_K_M GGUF) is bit-for-bit reproducible across 100 runs on the same machine. Built with uv, marimo, and PyTorch.

A framework for few-shot evaluation of language models.

A specialized ~500M parameter Arabic Small Language Model (SLM) paired with a decoupled micro-acoustic engine (244M Whisper-Small Arabic STT + Saudi-Native VITS TTS). Built for a single NVIDIA RTX 3070 (8 GB VRAM) under WSL2.

No project description available.