ML Engineer building GenAI and ML platforms at scale.
Currently at Singapore's largest grocery retailer, where I build LLM platforms, improve recommendation systems, and computer vision solutions for creatives.
- Systems programming in C and Rust for Apple Silicon (inference kernels, thermal control, SMC/IOKit)
- Local and on-device LLM inference: Metal kernels, GGUF, MLX, benchmarking real hardware limits
- AI developer tooling: MCP servers, CLIs, and a local-first memory graph
| Project | Description | Tech |
|---|---|---|
| atlassian-cli | Unified CLI for Jira, Confluence, Bitbucket & JSM. Bulk ops, dry-run, JSON/CSV/YAML output, multi-instance profiles. Docs | Rust |
| surge | LLM inference engine for the Mac Studio M3 Ultra. Limiter-aware pacing scheduler, byte-exact Metal decode, no dependencies beyond macOS. Work in progress | C, Metal |
| fanpro | Fan control and thermal monitoring for Apple Silicon. CLI, TUI and root daemon, zero dependencies | C, IOKit |
| speedlog | Internet speed monitor. One bash script, one HTML file, no Docker, no database | Bash, HTML |
| batteryconsole | Logitech MX device battery levels on macOS | Rust |
| logi_mx_auto_switch | Make an MX Master follow the MX Keys across Macs via HID++ ChangeHost, no Logitech software | Python |
| Project | Description | Tech |
|---|---|---|
| parsnip | Local-first memory graph for AI assistants. Single binary, entities/relations/observations, 5 search modes, cross-project queries. Site | Rust, redb, tantivy, MCP |
| gemini-mcp-rust | MCP server for Google's Gemini API | Rust, MCP |
| reddit-mcp-server | Read-only Reddit MCP server over app-only OAuth | Rust, MCP |
| skills | Custom skills for the Claude Code CLI | Markdown |
| Project | Description | Tech |
|---|---|---|
| llm-benchmark | Local LLM benchmark suite: 26 prompts, 6 categories, programmatic plus LLM-as-judge scoring | Python |
| bengali-ocr-finetune | Bengali OCR fine-tuning on Apple Silicon with mlx-vlm | Python, MLX |
| Project | Description | Tech |
|---|---|---|
| saas_template | Cloudflare-first SaaS starter. Opinionated, SEO-first, swappable. Demo | Next.js 16, D1, Better Auth, Stripe |
atlassian-cli: 28 stars, 7,217 GitHub release downloads, 541 crates.io downloads (as of 27 Aug 2026).
Merged: App-Store-Connect-CLI (submission validation, localization updates, price point filtering, stale review handling) and aws-codecommit-devops-model.
Open: psd-tools (drop shadow and outer glow layer effect rendering), whatsapp-mcp (context-aware whatsmeow API), HistoryHound (stdio transport, Chrome profile detection).
Systems: Rust, C, Metal, IOKit ML/AI: Python, PyTorch, MLX, GGUF, llama.cpp, Computer Vision, NLP AI tooling: Model Context Protocol (MCP), Claude Code Web: TypeScript, Next.js, Cloudflare Workers/D1 Cloud: GCP (Vertex AI, BigQuery, Cloud Functions), AWS (SageMaker, Lambda), Kubernetes, Docker
- macOS clamps my M3 Ultra's GPU to 338 MHz before the fans even try
- Kimi-Linear ran a real 1M context on my Mac Studio
- The two numbers that decide local LLMs: 100 tokens/sec and 1M context
- Local LLM Benchmark: Gemma 4 vs Qwen 3.5
- Serving ML Models Serverlessly (AWS UG Malaysia)
- Practical Introduction to NLP
- Website: omarshabab.com
- LinkedIn: /in/omar16100