leetscuda.com

Live at https://www.leetscuda.com

AI Infrastructure Interview Prep Platform — practice CUDA/GPU programming, ML systems, Kubernetes for AI, and systems foundations through real coding problems in your browser.

Targeted at senior/staff-level interviews at GPU-cloud and AI infrastructure companies (Nebius, Groq, Crusoe, Lambda Labs, Anyscale, etc.).

---

What it is

LeetsCUDA is a LeetCode-style platform where every problem involves something you'd actually encounter building AI infrastructure:

• Writing CUDA kernels (vector-add, matrix-transpose, parallel reduction)
• Implementing ML inference primitives (KV cache, continuous batching, Flash Attention)
• Writing Kubernetes manifests for GPU workloads (vLLM deployments, HPA with custom metrics)
• Systems analysis (roofline model, PCIe bandwidth, false sharing)

Problems are written in MDX, compiled and executed on distributed judge nodes, and presented in a split-pane editor powered by Monaco (the VS Code engine).

---

Quick start

Prerequisites

• Node.js >= 20
• pnpm >= 8
• PostgreSQL (local or Docker)
• Redis (local or Docker)

1. Clone and install

git clone https://github.com/isztld/leetscuda.com.git
cd leetscuda
pnpm install

2. Configure environment

cp apps/web/.env.example apps/web/.env

Edit apps/web/.env and fill in:

NEXTAUTH_SECRET=<generate with: openssl rand -base64 32>
NEXTAUTH_URL=http://localhost:3000

GITHUB_CLIENT_ID=
GITHUB_CLIENT_SECRET=

GOOGLE_CLIENT_ID=
GOOGLE_CLIENT_SECRET=

DATABASE_URL=postgresql://postgres:postgres@localhost:5432/leetscuda
REDIS_URL=redis://localhost:6379

3. Start Postgres and Redis

docker run -d --name leetscuda-pg \
  -e POSTGRES_PASSWORD=postgres \
  -e POSTGRES_DB=leetscuda \
  -p 5432:5432 postgres:16

docker run -d --name leetscuda-redis \
  -p 6379:6379 redis:7

4. Migrate, seed, and sync the database

pnpm --filter @leetscuda/web db:migrate   # apply schema migrations
pnpm --filter @leetscuda/web db:seed      # seed tracks + roadmap nodes
pnpm --filter @leetscuda/web db:sync      # sync problems from MDX files
pnpm --filter @leetscuda/web theory:sync  # sync theory from MDX files

db:sync and theory:sync must run after db:seed because problems reference track IDs created by the seed.

5. Start the dev server

pnpm dev

App runs at http://localhost:3000.

---

OAuth setup

GitHub

1. Go to https://github.com/settings/developers → New OAuth App
2. Set Authorization callback URL to http://localhost:3000/api/auth/callback/github
3. Copy Client ID and Client Secret into .env

Google

1. Go to https://console.cloud.google.com → APIs & Services → Credentials → Create OAuth client
2. Application type: Web application
3. Add authorized redirect URI: http://localhost:3000/api/auth/callback/google
4. Copy Client ID and Client Secret into .env

---

Repository structure

leetscuda/
├── apps/
│   ├── web/                    Next.js app (App Router)
│   │   ├── app/
│   │   │   ├── (auth)/         Sign-in, username setup pages
│   │   │   ├── (platform)/     Roadmap, problems list, problem detail
│   │   │   └── api/judge/      Judge HTTP API (poll, result, recover)
│   │   ├── lib/                Prisma client, Redis, auth, streak helpers
│   │   ├── scripts/            CLI scripts (create-judge-token, db-sync, theory-sync)
│   │   ├── server/
│   │   │   └── routers/        tRPC routers (user, roadmap, problems, submissions, profile)
│   │   └── prisma/
│   │       ├── schema.prisma
│   │       ├── seed.ts         Reads learning/*/track.mdx to seed tracks + nodes
│   │       └── migrations/
│   ├── judge-cuda-cpp/         C++/CUDA judge worker
│   │   └── ARCHITECTURE.md     Implementation reference
│   └── judge-k8s/              Kubernetes manifest validation judge
│       └── ARCHITECTURE.md     Implementation reference
├── docs/
│   ├── architecture.md         System architecture, DB schema, tRPC, auth
│   ├── judge.md                Judge deployment and operations guide
│   └── content.md              Problem format and content contribution guide
└── learning/                   All content (open for community PRs)
    ├── cuda-core/              CUDA & GPU Programming track
    │   ├── track.mdx           Track metadata + roadmap node definitions
    │   ├── problems/           vector-add, matrix-transpose, matrix-multiply, reduce-sum
    │   └── theory/             cuda-intro, cuda-threads, cuda-memory, cuda-streams
    ├── ml-systems/             ML Systems & Inference track
    ├── kubernetes-ai/          Kubernetes for AI track
    └── foundations/            Foundations track

---

Development commands

# Web app
pnpm dev                                    # start Next.js dev server
pnpm --filter @leetscuda/web lint
pnpm --filter @leetscuda/web build

# Judge workers
cd apps/judge-cuda-cpp && pnpm dev          # C++/CUDA judge (watch mode)
cd apps/judge-k8s && pnpm dev              # K8s judge (watch mode)

# Database
pnpm --filter @leetscuda/web db:generate    # regenerate Prisma client after schema changes
pnpm --filter @leetscuda/web db:migrate     # apply pending migrations
pnpm --filter @leetscuda/web db:seed        # seed tracks + roadmap nodes from learning/*/track.mdx
pnpm --filter @leetscuda/web db:sync        # sync problems from MDX files
pnpm --filter @leetscuda/web theory:sync  # sync theory from MDX files
pnpm --filter @leetscuda/web db:studio      # open Prisma Studio

# Token management
pnpm --filter @leetscuda/web judge:token --name <name> --capabilities <csv>

---

Contributing content

Problems, theory articles, and track definitions are open for community PRs — this is a core strength of the platform. See CONTRIBUTING.md for the full guide.

---

Documentation

• docs/architecture.md — Submission flow, tech stack, database schema, tRPC routers, authentication, environment variables
• docs/judge.md — Setting up and operating judge nodes (CPU, GPU, K8s)
• docs/content.md — Problem and theory content format reference
• CONTRIBUTING.md — Full contribution guide for content authors