WK 1 LIVE · CUDA, A→Z

Learning NVIDIA in public.

A 12-week curriculum to master NVIDIA's stack on a $4,700 DGX Spark. One week, one chapter, one hands-on project. Read along, build along, or fork the whole thing. New drops every Sunday.

★ NOW PLAYING · WK 01

What & Why of CUDA

The 17-year origin story, the three-layer stack, the DGX-vs-Mac comparison everyone needs, and a Sunday capstone where we fine-tune Llama on US real estate data.
UP NEXT · WK 02

Inference at Scale

vLLM, TensorRT-LLM, continuous batching. The week we go from "works on my desk" to "could host this for 100 paying users."

The 12-week program

~7 days each. ~1–2 hours per day. One hands-on project per week.

WK 01
What & Why of CUDA
Origin story · the moat · DGX vs Mac · pretenders · landscape · your bet · fine-tune capstone
LIVE WK 02
Inference at Scale
vLLM · TensorRT-LLM · continuous batching · paged attention
SOON WK 03
Fine-Tuning Deep Dive
LoRA · QLoRA · DPO · SFT vs RLHF · evals that matter
SOON WK 04
Quantization & the FP4 Revolution
FP16 · BF16 · INT8 · FP4 · AWQ · GPTQ · Marlin kernels
SOON WK 05
Local Agents & Tool Use
Function calling · MCP · tool routing · multi-step reasoning
SOON WK 06
RAG Done Right
Embeddings · vector DBs · rerankers · evaluation · the failure modes
SOON WK 07
Multi-Modal
Stable Diffusion · Whisper · F5-TTS · LLaVA · vision agents
SOON WK 08
Real-Time AI
Streaming · voice agents · sub-second latency engineering
SOON WK 09
Custom CUDA via Triton
Your first real GPU kernel — in Python, not C++
SOON WK 10
Production Serving
NVIDIA Triton Inference Server · monitoring · scaling · cost
SOON WK 11
Training Your Own Architecture
Beyond fine-tuning · novel models · the deep cut (optional)
SOON WK 12
Capstone — Ship a Real Product
Twelve weeks compounded into one shipped thing
SOON

What this is

In 2026 a $4,700 NVIDIA DGX Spark sat next to my couch. I'm a non-engineer founder who vibe-codes with Claude Code and ships local-first AI products at LocalsOnly.AI. I bought the box knowing it was important and not really knowing why. So I'm doing what I always do: learn it in the open, week by week, with a real project at the end of each one.

No prerequisites, no PhD. Each week is built so an operator with a terminal and a working SSH connection can follow along. Every chapter ends with a hands-on project — usable code, useful screenshots, something that runs on your own hardware. By Week 12 you have twelve real things you built, and the kind of CUDA fluency that makes you dangerous in any AI conversation.

Cadence
Weekly · Sunday drops
Length
12 weeks
Hardware
NVIDIA DGX Spark · $4,700
Cost to follow
$0 (read-along free)

Want to follow along? Bookmark this page — new chapter every Sunday. Want to run the labs yourself? You'll need a DGX Spark or any NVIDIA GPU with ≥16 GB of memory and CUDA 12+.

Start Week 1 → Read the Apple comparison