Data, models & applied AI, distilled into useful references.
I'm Omar — a data scientist and ML engineer with 26+ years across credit, telecom and EDA. This site is where I keep my cheatsheets and field notes on machine learning, LLMs, agents and the tools I actually use.
Two libraries, one mission
Quick references for fast lookup, long-form articles for deeper understanding.
Cheatsheets
Concise reference guides for the tools and libraries that show up in real data-science and ML work. Skim them, bookmark them, print them.
Browse cheatsheets →Blog & Articles
In-depth articles, capability tours and field notes on emerging AI models and tooling — things I learned from actually running them, not just reading about them.
Read the blog →Latest from the site
A few of the newest entries from both libraries.
🕸️ LightRAG — fast graph-RAG with GPT-5 + Qdrant, hands-on
Graph-RAG without GraphRAG's bill: knowledge-graph indexing plus a dual-level retrieval paradigm, run end-to-end on OpenAI gpt-5-mini / gpt-5.5 and a Qdrant vector store — the seven core ideas, the knowledge graph it builds, and the real setup issues and fixes.
Article · June 2026🕸️ LinearRAG — relation-free GraphRAG, hands-on a 4090
GraphRAG without relation extraction: a Tri-Graph built with zero LLM tokens, single-pass multi-hop retrieval via semantic bridging and Personalized PageRank, animated walkthroughs, and an end-to-end 4090 run where gpt-5-mini scored a perfect 12/12.
Article · June 2026🎨 Boogu-Image — the Edit and Turbo models, hands-on a 4090
The Apache-2.0 10B unified image model on a 24 GB 4090: instruction-based photo editing and a 4-step Turbo text-to-image — install, prompt/config galleries, and the honest blur / OOM / group-offload / identity-CFG story.
Article · June 2026🔭 Lens — quantizing the DiT to run 1440 without offload on a 4090
Microsoft's 3.8B text-to-image model with a 4-bit GPT-OSS encoder: the no-offload OOM wall and the FP8 DiT fix that unlocks 1440, with a category gallery and honest quality notes.
Article · May 2026⚡ Doubling Qwopus 3.6 on a single RTX 4090 — Multi-Token Prediction with llama.cpp
~1.98× average speedup (peak 2.21×) on Qwopus 3.6-27B-MTP. Why Ollama refused to load it, and the working llama.cpp install path end-to-end.
Cheatsheet · New📊 bupaR — Business Process Analytics in R
Event logs, process maps, performance metrics, filtering, and the bupaverse ecosystem in one quick reference.
Cheatsheet🤖 OpenAI Agents SDK
Core concepts, agent creation, tools, guardrails, handoffs, and tracing for building AI agent applications.
About me
I'm a seasoned data scientist and ML engineer with 26+ years of experience. I've contributed to Tribal Credit, Reveel, TA Telecom, Mentor Graphics and IBM — across AI-driven underwriting, algorithm development, analytics and software-testing leadership.
I write and build around: