The Architect
A teaching game that turns 1,600 real production AI case studies into playable architectural decisions. You don't read “use a reranker.” You ship without one, watch your Quality meter crater, and feel why.
- Next.js (App Router)
- TypeScript
- Tailwind
- Zustand
- Supabase
- Claude API (RAG)
The idea
Most ML/LLM system-design learning is static: blog posts and flashcards you read and forget. The Architect drops you into architectural decisions real companies actually faced in production — DoorDash search latency, a support bot hallucinating refund policy — lets you decide, and moves three meters in real time: Cost, Latency, Quality.
Then comes the reveal: what the real company (Netflix, Ramp, DoorDash) actually built, why, and the tradeoffs they lived with. That reveal is the dopamine and the learning at once.
Grounded in real systems
Every scenario comes from a real production system, not an invented one. The ZenML LLMOps Database distils 1,600+ engineering blog posts into a structured dataset — one company, one architecture, the decisions and tradeoffs they actually hit.
That dataset is the spine of the game: each case study becomes a scenario, each recurring pattern becomes an unlock. The shape of the data is the shape of the play.
Design
Terminal-meets-comic: a dark workspace with real-looking architecture diagrams and logs on one side, illustrated characters arguing different directions on the other, three meters ticking at the bottom, decisions styled as terminal commands.
The first playable track is Evals — one deep, polished loop rather than a broad, shallow survey, with the cross-cutting patterns (retrieval, routing, guardrails) layering in from there.
Why it matters
It's the playable form of the AI-engineering work I do every day: retrieval, evals, routing, guardrails, agents, and the cost/latency/quality tradeoffs between them. Turning those tradeoffs into something you can feel — not just read — is the clearest demonstration of understanding them in depth.