Systems and full-stack software engineer contributing directly to industry-standard machine learning runtimes, silicon kernels, distributed inference engines, and scalable developer ecosystems.
 
|
 
|
 
|
Active core contributor to tier-1 AI accelerators, distributed inference runtimes, and developer frameworks:
| Organization / Project | Focus Area | Key Contributions & Pull Requests | Status |
|---|---|---|---|
 Google |
Gemini Developer Platform |
#1436: fix(link-checker): exclude fenced code blocks from anchor extraction (fixes #1435) |
 |
 Tenstorrent |
Silicon Kernels & CCL Runtimes |
#61500: fix(ccl): credit alternate route in multicast all-gather on even rings (fixes #61484)#55598: fix(moreh): prevent non-finite NaN in norm_backward for zero-norm slices#55590: fix(sfpu): Chebyshev polynomial bridge for lgamma precision (<4 ULP) |
  |
 vLLM Project |
Distributed Inference Engine |
#59445: fix: prevent path traversal outside cache directory in FilesystemResolver (fixes #59270)#60649: fix(benchmarks): support boolean filter values in sweep plot CLI |
 |
 Qdrant |
Vector Embedding Runtimes |
#717: Normalize dynamic tokenizer padding for ONNX embedding pipelines (co-author) |
 |
 Hugging Face |
Transformers Framework |
#48549: feat(hf_argparser): support dict-typed dataclass fields from CLI (fixes #48030) |
 |
 LiteLLM |
AI Gateway & Proxy |
#47191: fix(transformations): raise on invalid completion responses for fireworks/ollama#40282: fix(streaming): preserve reasoning tokens and usage in chunk builder#41603: fix(proxy): resolve model_group_alias in vertex live setup extraction |
 |
 LibreChat |
Full-Stack AI Architecture |
#16223: Resolve file search indexing when attaching assets directly from files panel |
 |
 OpenAI |
SDK & Workload Identity |
#4017: fix(auth): preserve response and prevent duplicate error wrapping in workload providers |
 |
- ⚙️ Low-Level Systems & ML Runtimes: Deep expertise in AI accelerators (Tenstorrent Wormhole/Blackhole), quantization kernels (FP8, MXFP8, Int8), and high-throughput LLM serving via vLLM.
- 🌐 Enterprise Open Source: Consistently landing high-precision fixes across industry-defining stacks including Google, vLLM, Tenstorrent, Qdrant, Hugging Face, and LiteLLM.
- 🚀 Full-Stack Architecture: Building high-concurrency microservices, distributed real-time messaging, and reactive applications using Python, TypeScript, C++, and Flutter.
- 🎯 Engineering Discipline: Focused on idiomatic code, minimal diffs, rigorous regression testing, and eliminating numerical instability in tensor operations.
| Project | Tech Stack | Overview |
|---|---|---|
| gemma-vf-agent | Python Gemma 4 DeepMind Pytest AST |
Deterministic Verification-First Agentic Software Engineering with Gemma 4 (Google DeepMind Kaggle Paper Track). |
| system-design-pro | Architecture Distributed Systems Capacity Sizing Agent Skill |
Production Distributed Architecture Blueprints, Capacity Sizing Engine & Universal Agent Skill for Antigravity, Claude Code & Cursor. |
| nbox | C++ Neural Video Real-Time Upscaler ONNX |
Subtle Neural Video Enhancer & Real-Time Upscaler (240p -> 2K/4K) with Zero CPU Load. |
| Libya_ID | Python FastAPI Reverse Engineering PostgreSQL |
High-speed real-time caller identification and indexing engine built for local Libyan telecom carriers. |
| Card SMS & Virtual Numbers Hub | TypeScript Node.js PostgreSQL Redis Docker |
High-throughput virtual SMS verification gateway featuring multi-provider failover, rate-limiting windows, and transactional delivery. |
|
|
|
|
|
|
I'm always open to discussing new opportunities, open-source initiatives, or innovative high-performance engineering:
- 🌐 Platform & Portfolio: nbox.ly
- 💼 GitHub: @mohmedmm
- 📧 Direct Email: alshaykh@nbox.ly
- 📍 Location: Tripoli, Libya
Designed & engineered with precision by Mohammed Alshyakh • Open for high-impact systems & ML collaborations! ⭐