Most system design "practice" grades you on whether you drew the right boxes cache here, load balancer there checked against a keyword list. That never sat right with me, because the actual skill being tested in a system design interview isn't "do you know the vocabulary?" it's "Can you reason about what happens to this system under load."
So SystemInPain doesn't grade the diagram. It runs it. There's a deterministic simulation engine underneath the same architecture and the same seed always produce the exact same result and it pushes real seeded traffic through whatever you build, then measures p99 latency, availability, cost, and throughput under both steady load and injected failure scenarios (traffic spikes, node outages, and bot floods).
honestly this looks great for actually seeing where things break. one thing that would help me a lot though is letting you replay a level and tweak just one variable like node count or arrival rate, basically see how small changes shift p99 without redoing everything from scratch