The AI Transformation Consultant
What works, what fails, and how to lead the engagement — 10 cited chapters from LLM fundamentals to boardroom strategy and an interview toolkit.
Open for Advisory & Collaboration · Cairo & California
Applied AI, graph retrieval & production ML, distilled from first principles.
I'm Omar — a data scientist and enterprise AI architect with 26+ years across fintech underwriting, telecom, EDA and distributed software. This site is my working library: six illustrated books, six free video courses, empirical GPU field notes, and 30+ engineering cheatsheets.
Career Pedigree //
HCLTech
CaaStle
Tribal Credit
Mentor Graphics
IBM
Lucent
01 // Published Monographs
Six books, illustrated & hands-on.
LLM Evaluation & Reliability Engineering
Evals, observability, and AgentOps for developers who own what happens in production — every chapter opens with a real incident and ships runnable Python.
The Camera Composition Playbook
Photography composition in 60 visual rules: weaker-vs-stronger frame comparisons, 10-minute field exercises, and when to deliberately break the rule.
Microservices for Python Developers
Follow ShopFlow from a single Django monolith to FastAPI microservices deployed across AWS and Google Cloud, paired with a free 4-module video course.
Advanced GraphRAG & Knowledge Agents
Beyond flat vector search: ontology construction, Microsoft GraphRAG & DRIFT, LightRAG, HippoRAG, and multi-hop knowledge agents across eight modules.
02 // Architecture of the Site
Four libraries, one engineering standard.
Quick-lookup syntax sheets for daily coding, empirical benchmarks from a local 24 GB GPU rig, structured video courses, and full-length books.
Library A · Reference
Engineering & ML Cheatsheets
Dense, printable multi-column command & API references with live command filtering, one-click copy, and A4 print layout — covering Scikit-learn, PyTorch, SQL, Docker, LangChain, PEFT, and AI agents.
Browse all 30+ reference sheets
→
Library B · Long-Form
GPU Field Notes & Blog
Empirical capability tours and architecture explainers written from actually running, quantizing, and benchmarking open-weight models on a single 24 GB RTX 4090.
Read the technical journal
→
Library C · Video Curriculum
Free Structured Video Courses
Self-contained 15–35 minute modules on AI strategy for consultants, reinforcement learning with Stable-Baselines3, GraphRAG Parts 1 & 2, LLM context & agent memory, and Python microservices.
Open the course syllabus
→
Library D · Print & Digital
Illustrated Technical Books
Friendly, visual, rigorously cited guides that pair plain-English mental models with runnable code, architecture diagrams, and companion YouTube courses.
Explore all six editions
→
03 // Recently Published
Field notes & benchmarks.
AI Agents · New
OpenWiki — the AI agent that writes your docs (and turns Gmail into a second brain)
Sandboxed docs-only repo documentation plus 100 real Gmail threads distilled into a terminal-queryable personal wiki, with SHA-256 snapshot guardrails.
Graph Databases
FalkorDB, explained — the graph database that's secretly linear algebra
How storing graphs as sparse matrices turns traversal into matrix multiplication, paired with native vector search for RAG and agent memory.
Style Atlas
12 Illustration Styles Worth Generating in Krea 2
Twelve master illustrator aesthetics reproduced locally on one 24 GB 4090 using pure visual-language cues rather than artist names.
Quantization
Krea 2 on a single RTX 4090 — quantizing a 12B image model with GGUF
Fitting a 12B diffusion transformer into ~10.6 GB under Pop!_OS: ComfyUI isolation, encoder gotchas, and a 16-style quality audit.
Document VLM
Unlimited-OCR, hands-on — long-horizon document parsing on a 4090
Near-exact English Markdown & tables vs 1890s historical Arabic pages, plus fixing runaway generation loops with image_mode=base.
GraphRAG
LightRAG — fast graph-RAG with GPT-5 + Qdrant, hands-on
Knowledge-graph indexing and dual-level retrieval without community-traversal bills, plus incremental updates and real setup fixes.
ICLR 2026
LinearRAG — relation-free GraphRAG, hands-on a 4090
Building a Tri-Graph with zero LLM tokens via NER & embeddings, then single-pass multi-hop retrieval with Personalized PageRank (12/12 accuracy).
Unified 10B
Boogu-Image — the Edit and Turbo models, hands-on a 4090
Instruction photo editing and 4-step Turbo generation on 24 GB VRAM: group offload, OOM walls, and identity-CFG experiments.
FP8 DiT
Lens — quantizing the DiT to run 1440 without offload on a 4090
Breaking past the 1024px bf16 OOM ceiling on Microsoft's 3.8B model + 4-bit GPT-OSS-20B encoder using on-GPU FP8 quantization.
1.98× Speedup
Doubling Qwopus 3.6 on a single RTX 4090 — Multi-Token Prediction
1.98× average throughput (2.21× peak on math) on a 27B GGUF model — why Ollama rejected it and the exact production llama-server recipe.
patients %>%
filter_activity_frequency(percentage = 0.8) %>%
process_map(performance(median, "days"))
filter_activity_frequency(percentage = 0.8) %>%
process_map(performance(median, "days"))
bupaR — Business Process Analytics & Process Mining
Event logs, frequency & performance process maps, trace filtering, dotted charts, and conformance workflows on one sheet.
agent = Agent(name="Triage",
handoffs=[billing_agent, tech_agent],
input_guardrails=[safety_check])
handoffs=[billing_agent, tech_agent],
input_guardrails=[safety_check])
OpenAI Agents SDK — Tools, Handoffs, Guardrails & Tracing
Complete multi-agent orchestration reference: Runner loops, structured outputs, function tools, handoffs, and trace instrumentation.
04 // Free Video Curriculum
Learn step by step — 49 modules.
AI for Business Consultants
Classical ML to LLMs, MCP, hyperscalers, EU AI Act, transformation failure modes, and a 15-question interview toolkit.
Context Management & Agent Memory in LLMs
Why models forget, SWE-Pruner, Context-Folding, Memory-as-Action, ACE playbooks, EverMemOS, MAGMA graphs, and CommitKV.
Advanced GraphRAG & Knowledge Agents (I & II)
From Microsoft GraphRAG, LightRAG & HippoRAG to LazyGraphRAG, temporal validity clocks, Graphiti/Zep, GNN-RAG & hypergraphs.