AI Systems Engineer based in Riyadh, SA.
I design agentic pipelines in pure Python — no framework bloat, no abstraction overhead.
Everything runs locally on AMD GPUs via ROCm: RDNA4 with llama.cpp, and two 2018-era MI50s with vLLM and custom engines.
Security isn't an afterthought: eJPT · eMAPT · eWPT-level web pentesting.
Running big models on old hardware: Qwen3.8-27B with a 200k-token context and 74 tok/s with speculative decoding, and DeepSeek-V4.1-Flash (552B MoE) at ~10 tok/s, both on two AMD MI50 16 GB cards from 2018. Custom HIP kernels, inference-engine ports and every number from a logged run.
-
Agentic AI
Multi-agent orchestration, pipeline design, local LLM inference, pure Python systems -
LLM Inference on AMD
vLLM · llama.cpp · custom HIP kernels for gfx906 · speculative decoding (DFlash2, MTP) · MoE expert offloading -
Infrastructure
Proxmox · OPNsense · LXC · WireGuard · Cloudflare Tunnel · multi-VLAN homelab -
Security
Web pentesting (eWPT-level) · network analysis · eJPT certified · eMAPT completed -
Systems & Tooling
ESP32-CAM · YOLO · ROCm/AMD · SQLite · Go · static-first web (no frameworks unless earned)
| Project | Description |
|---|---|
| mi50-vllm-lab | Qwen3.8-27B on 2× AMD MI50: 200k-token agent context, 74 tok/s with speculative decoding, five custom HIP kernels |
| DeepSeek on MI50 | Project Maya's DeepSeek-V4.1-Flash engine ported to gfx906: ~10 tok/s on 32 GB of VRAM, plus a page-cache fix for decode freezes |
| Agentic AI Architect | Framework-free multi-agent orchestration in pure Python |
| agent-fitness | Phase-based reasoning pipeline: Analyze → Think → Summarize → Decide |
| AI Agent Subtitle Downloader | 9-agent LangGraph pipeline · Sonarr + SubDL APIs |
blog.sbm-labs.com: write-ups in Arabic and English, including the dual-MI50 series on running LLMs on old AMD GPUs.
Design for clarity. Build for control. Ship without the bloat.