One machine on a desk.Building LLMs, running them, and teaching what it takes.
danny22.com runs on a single NVIDIA DGX Spark. The course material I write while serving, training and optimizing models here, and the system that keeps this site running, are published as labs.
Labs
Each lab is a course made of theory, hands-on work and code that actually runs. Every result is measured on this machine. The course material is written in Korean.
-
Data Culture Lab
A common skeleton for solving problems with data: ask where a problem came from, decide whether it is worth solving, fix how the result will be judged before building, turn results into decisions and look back. Cases come from six subsidiaries of a fictional holding company, and every other lab follows this skeleton.
In progress11 parts, 52 units -
Agent Lab
Learn to build agentic AI agents that plan, use tools, check results and carry multi-step work through to the end. Covers planning and tool calling, memory, multi-agent collaboration, evaluation and guardrails.
Coming later -
LLM Lab
Learn to serve an open LLM behind an API, make it fast and small, implement its insides yourself, and train it for the job with LoRA and full fine-tuning. An advanced part builds a small model from scratch.
In progress8 parts, 57 units -
WWW Lab
How to grow this site: search visibility, traffic analysis and talking with visitors, using its real data.
Planned6 parts, 27 units -
Platform Lab
Learn to turn one machine into a platform you can reach from anywhere (VPN), publish safely (tunnel, reverse proxy, login), restore when it breaks (backups) and that tells you first (regression tests, alerts), and hand its operation to an AI coding agent. This site is the result.
First draft11 parts, 39 units
How this site runs
The page you are reading comes from a single machine along this path. The machine opens no inbound ports.
- InternetYour browser looks up www.danny22.com.
- DNS and security serviceResolves the name and encrypts the connection.
- TunnelRequests arrive over a connection the machine opened outward.
- Reverse proxyRoutes requests by address and checks login where needed.
- ServicesThis page, the labs and the login service answer.
An AI coding agent operates the site. It runs regression tests before and after every change, backs up every night, and sends an alert when something breaks.
Changes that are hard to undo, such as DNS, deployments and deletions, need a person's approval.
Contact
Questions about the courses, collaboration or anything in the labs are welcome. I will reply to the email you leave.