System-design interview trainer

🏗️ CloudArchitect-Sim

A drag-and-drop, LLM-graded canvas for practicing system-design interviews. Draw the architecture, an interviewer scores it against a rubric, asks one follow-up, you revise — up to three rounds.

← Back to the game

The core loop

Nothing reaches the LLM until the diagram survives a deterministic gate — that's what keeps grading cheap and scores reproducible.

Canvas

Drag nodes, wire connections (HTTP / gRPC / WS / async)

JSON topology

Graph serialized: nodes, edges, config, region, notes

Deterministic gate

Code-only checks — a hard fail short-circuits, no LLM spend

LLM interviewer

Difficulty persona scores 5 categories + one follow-up

Revise

Answer in text or redraw; re-evaluate

↻ loops back to 01 — capped at 3 rounds per level

Architecture

Client

frontend/

Next.js 14 · React Flow · Tailwind

Canvas & palette — provider-agnostic components
Cost indicator — live monthly estimate
Scorecard — 5-axis radar, deterministic issues
Stress sim overlay — RPS particles, bottleneck node
Progress panel — pass rate, weakest category, debrief

Server

backend/

FastAPI · Firestore · offline mock mode

deterministic.py — mandatory components, anti-patterns
prompt_builder.py / llm_client.py — persona prompts
cache.py — topology-hash response cache
conversation.py — follow-up state, LRU-evicted
ratelimit.py — per-minute limit + daily quota + spend cap
db.py — player history & share links, Firestore-backed

Campaign levels

#	Level	Requires	Scope
01	Webhook Router	API Gateway + a DB — queue/cache/warehouse is over-engineering	v1
02	Automated Media Processing Pipeline	Queue/Pub-Sub + Object Storage — async decoupling	v1
03	High-Frequency Trading Bot	In-memory cache + load balancer — no SPOF	v1
04	Enterprise Big Data Analytics	Data warehouse + queue/pub-sub — ingestion pipeline	v1
05	IoT Telemetry Platform	Queue/Pub-Sub + time-series DB — buffered ingestion	v2
06	Recommendation Engine	Graph DB + in-memory cache — graph serving + cache	v2
07	RAG-Based Search	Vector DB + object storage — vector serving	v2
08	Always-On Payments API	Load balancer + relational DB — no compute SPOF, replicated DB	v2
09	Multi-Region Disaster Recovery	Compute + data in ≥2 regions — regional DR	v3
10	Globally Resilient Ledger (5 nines)	Multi-region + replicated DB + no SPOF — consistency trade-off	v3

Feature ledger

v1 · MVP — core loop

Canvas & palette — drag/click placement, typed connections
Live cost indicator — monthly estimate as you build
Scorecard — 5-axis scores, issues, interviewer comment
Follow-up loop — up to 3 rounds, text or redraw
Reference solution — unlocked on a pass
Offline mock interviewer — fully playable with no API key

v2 — depth

Component config — DB/cache replicas, cache TTL, multi-AZ
Expanded DB palette — time-series, graph, vector stores
Stress simulation — target RPS, live bottleneck highlighting
Save & Share — read-only link to a scored design
Levels 5–8 — same acceptance-criteria format

v3 — breadth

Multi-cloud skin — AWS / GCP / Azure / generic naming, one logic layer
Cross-region scenarios — per-node region, DR/HA structural checks
Mock interview mode — open-ended scenario, 15-minute timer
Custom scenario — paste a real prompt, LLM structures the brief

Cost & production guardrails

Response cache

topology-hash

identical designs skip the LLM entirely

Rate limit

per-minute + daily quota

CLOUDARCH_RATE_PER_MIN / _DAILY_QUOTA

Global spend cap

daily budget

over budget → auto-degrades to offline interviewer

Session store

bounded, LRU

idle conversations evicted on TTL

Player data

Firestore

history & shared designs, keyed by a browser-local ID — no account/login

Hosting

GCP Cloud Run

serverless containers, europe-west1, scale-to-zero

Test coverage

68 + 5

backend pytest + frontend vitest, both build clean

Golden-set eval

17 fixtures × 3

real-LLM regression harness, run on prompt changes

← Back to the game