AI engineering, built from scratch
Build real AI systems from scratch
Clone a repo, implement a real system stage by stage, and get instant, deterministic pass/fail feedback. No toy models — real GPT-2 weights, real specs, real numbers. All in TypeScript.
theinference — test --watch
✓ PASS Load the GPT-2 weights 7ms✓ PASS Implement causal self-attention 210ms✓ PASS Match GPT-2 logits 1.2sStage 3/3 complete ✓ — synced to the siteHow it works
A loop that actually teaches
The fast, satisfying inner loop is the whole point. You learn by building the real thing and watching it pass.
Step 1
Clone the starter
One command scaffolds the project with pinned dependencies. No GPU, no API keys, no setup drift.
Step 2
Implement a stage
Each stage is a real piece of the system — attention, a tokenizer, a retriever. You write the actual logic.
Step 3
Get instant feedback
A local grader runs in milliseconds and tells you exactly what is off — within tolerance, byte-exact, or recall@k.
The catalog
Start with these
Build your own LLM today
The anchor course is free, runs entirely on your laptop, and you can clear the first stage in under ten minutes.