◆ SafeZone AI Learn

theory → playground → your own GPU

Learn the AI stack by watching it work.

Every lesson pairs real theory with a live experiment: draw a maze and race A* against Dijkstra, train an agent that teaches itself Pac-Man, distil and interrogate an LLM. Nothing is a video — everything runs, on your own machine, nothing hidden.

Start with the theory → Open the playground

Pick a track

One platform, growing toward the full ML · AI · DevOps stack. The first track is live.

Reinforcement Learning

Search → RL → agentic LLMs: six lessons that each reuse the one below, from a maze you draw to a Dean that teaches a model to mastery.

6 lessons · liveStart →
in development

ML Fundamentals

Regression to gradient boosting — watch models fit, overfit and generalise on data you can poke.

coming
in development

LLM Engineering

Tokenisers, attention, RAG and evals — the stack behind every AI product, opened up live.

coming
in development

DevOps

Containers, CI/CD, infrastructure as code and cloud deployment (AWS · Azure · GCP) — ship what you build, the way this platform ships itself.

coming

Learn

Every algorithm and method, explained properly — the idea, the math, pseudocode, and when to reach for it.

Read the theory →

Play

The practical half. Draw a maze, race the algorithms, train an agent, teach an LLM — and see every result animate live.

Open the playground →

Your machine

The playground computes on your GPU via a small worker you run locally. Your models and API keys never leave it.

How it works →

Inside the RL track

Each lesson reuses the one below it. Follow the path top to bottom, or jump anywhere.

  1. 1

    Grid maze worker

    Classic pathfinding on a maze you draw — BFS, DFS, Dijkstra, A*, greedy and bidirectional search, racing side by side.

    Theory · Try it
  2. 2

    Real streets worker

    The very same searches, now racing over a live OpenStreetMap road graph. A* vs Dijkstra, made tangible on real cities.

    Theory · Try it
  3. 3

    Deep RL GPU

    An agent learns Pac-Man & Pong straight from the pixels — the DQN family and the policy-gradient family (REINFORCE, A2C, PPO).

    Theory · Try it
  4. 4

    Teach an LLM GPU

    Distil a large teacher into a small student model — supervised fine-tuning, then DPO, the RL-from-AI-feedback step.

    Theory · Try it
  5. 5

    Reasoning Lab GPU

    Wire prompted agents into generate → critique → refine loops — the Self-Refine / Reflexion pattern, built by hand.

    Theory · Try it
  6. 6

    University GPU

    A Dean agent designs a curriculum and teaches a model each subject to mastery — solo, or a competing cohort of students.

    Theory · Try it

Start at the bottom of the ladder.

No account. The theory is free and instant; the playground runs on a worker you download and run.

Read: Pathfinding → Jump into the playground