Courses
Rebuild the measurements from the blog yourself. Every course is video plus the notebooks used on screen, and at least the first 3 lectures of each are free.
Tracks:BUILDwrite it from scratch →HANDS-ONuse the real tools →SHIPrun it in production
Highlights
All courses
15 coursesAttention and Transformers from Scratch
From RNN Seq2Seq via Bahdanau and Luong attention to a full Transformer, built and trained in PyTorch
Build · 28 lectures · 3h 24m · first 3 free
DDPM and DDIM from Scratch
Derive and implement DDPM and fast DDIM sampling in PyTorch, from the paper to working code on MNIST
Build · 19 lectures · 2h 16m · first 3 free
Latent Diffusion to SANA from Scratch
Stable Diffusion, VAE and CFG, Diffusion Transformers, PixArt-α and SANA, each implemented in PyTorch
Build · 23 lectures · 2h 37m · first 3 free
Neural Network Quantization Hands-On: PyTorch and ONNX
Number formats, quantization math, dynamic/static/QAT, ResNet and YOLO on ONNX Runtime, all measured
Hands-On · 41 lectures · 5h 39m · first 3 free
LLM Quantization and Compression Hands-On
GPTQ, AWQ, bitsandbytes, GGUF, QLoRA, pruning and distillation: fit LLMs into the memory you have
Hands-On · 24 lectures · 3h 19m · first 3 free
GraphRAG Hands-On: Knowledge Graphs for RAG
Why vector RAG fails, what a knowledge graph fixes, how to index and query with Microsoft GraphRAG
Hands-On · 17 lectures · 1h 47m · first 3 free
GraphRAG in Production
Real documents, production indexing and cost, LLM-judge evaluation, provenance and a GraphRAG agent
Ship · 19 lectures · 2h 8m · first 3 free
On-Device LLMs Hands-On: Prepare and Compress
Compression theory, choosing a small model and converting it with llama.cpp, MLC LLM and ONNX Runtime
Hands-On · 19 lectures · 2h 33m · first 3 free
On-Device Stable Diffusion Hands-On: Profile and Export
Pipeline profiling, few-step LCM and SDXL Turbo, ExecuTorch export with INT8 calibration, Qualcomm QNN
Hands-On · 18 lectures · 2h 24m · first 3 free
Ship an LLM in an Android App
Kotlin with Compose and MediaPipe, React Native with llama.rn, hardware acceleration and NPU verification
Ship · 14 lectures · 1h 42m · first 3 free
Ship Stable Diffusion in an Android App
ONNX Runtime Mobile in Kotlin and React Native, acceleration, an offline assistant capstone and a benchmark
Ship · 14 lectures · 1h 55m · first 3 free
RAG Evaluation in Production: Trustworthy Agents
Score a RAG system with Ragas and LLM judges, diagnose its retrieval, evaluate the agent
Ship · 18 lectures · 2h 45m · first 3 free
Build a Jev-Style Decision Model from Scratch
An LLM that returns calibrated probabilities, not text: architecture, PyTorch build, training, serving
Build · 23 lectures · 1h 58m · all lectures free
Claude Code Hands-On: CLAUDE.md, Hooks, Skills, Subagents
Drive Claude Code in a real repo: sessions, permissions, CLAUDE.md, hooks, skills and subagents, recorded live
Hands-On · 40 lectures · 1h 58m · first 3 free
Claude Code in Production: MCP, Headless, Agent SDK
Connect tools with MCP, build your own server, run Claude Code headless in scripts and CI, and program it with the Agent SDK
Ship · 16 lectures · 42m · first 3 free
What learners say
Real reviews from the instructor's courses, translated from Korean.
“Thanks to the well-structured curriculum and the instructor's clear explanations, I successfully transitioned into an LLMOps role.”
“I knew nothing about LLMs beyond prompts, but after finishing the course I applied it at work — and now I can pick up new techniques as they come out.”
“It covers things that are hard to learn from books or papers, delivered systematically, with hands-on labs built to make the concepts click. Highly recommended.”
“What I valued most: it goes beyond theory into prompt design and LLM applications you can actually use at work, all practice-driven.”
“RAG and prompting explained through the papers themselves. Picking good papers is hard, understanding them is harder — walking through them with hands-on practice made the concepts easy to grasp.”
“A solid introduction to Transformers — but a cut above any of the video lectures you can find on YouTube or elsewhere on the web. Thank you.”
“I wandered all over trying to learn Diffusion, stumbled onto this course, and finally understood it completely.”
Bundles
Compared with buying each course on its own
AI Engineering Bundle · 13 courses
34h 25m · Every AI engineering course on sotaaz — lifetime access, one payment. One by one: $628.
Claude Code Bundle · 2 courses
2h 40m · Both Claude Code courses, from the daily loop to MCP, CI and the Agent SDK — lifetime access, one payment. One by one: $68.
Free code from the posts
The scripts and raw results behind the measurements. Free with an account.
KV Cache Benchmark Harness
Free- llama-bench harness: Prefill, generation, perplexity and decode at 8K/32K/64K depth
- llama-server harness: Concurrent slots, TTFT and throughput, VRAM at ready and peak
- Raw results: Per-request JSON and server timings from both posts
- Four pitfalls: CPU fallback, the 64KB SSE buffer, VRAM sampling, token counting
GPT-2 to the Speedrun: Code and Raw Results
Free- Notebook: Reproduces every number and figure in the posts from the shipped logs; flags to download the data and train
- Training code: GPT-2 (124M) with one switch per change, a fixed-budget trainer and the GPT-2 logit check
- Raw results: Every run's step-by-step log, including the attempt that was killed
- Pre-registered plan: Questions and decision rules committed before the runs, with commit times
For non-developers, on Udemy
How ChatGPT Actually Works: LLM Principles Without Code
English edition — available now
How ChatGPT actually works, explained without code — for non-developers who want to understand what they are using.
View on Udemy →How ChatGPT Actually Works — Korean Edition
코드 없이 이해하는 ChatGPT 원리 — 비개발자를 위한 LLM 입문
The original Korean edition of the LLM fundamentals course.
View on Udemy →Prompt Design That Survives Model Updates — Practical AI, Principled
원리를 알고 써먹는 ChatGPT 실전 — 비개발자를 위한 프롬프트 설계와 AI 활용법
Prompt design grounded in principles, not templates — practical AI skills that survive model updates.
View on Udemy →AI in Practice for Data Analysts
분석가를 위한 AI 실무 — EDA·SQL·보고서 자동화와 검증
Where AI actually fits in an analyst's workflow — AI-assisted EDA, text-to-SQL you can trust, document Q&A, and verifying every number AI gives you. Measured error rates, not tips.
New courses by email
One weekly newsletter. New courses and the measurements behind them go in it.