An open-source, enterprise-grade coding interview platform inspired by Meta's 2025 AI-Enabled Coding Interview format — multi-file project workspace, governed AI assistant, and Checkpoint-driven scoring in a single seamless experience.
In late 2025, Meta became the first FAANG company to officially adopt AI-assisted coding interviews, reflecting the reality that 82% of developers already use AI tools at work. The interview is no longer "how fast can you code alone" but "how well do you think, prompt, and collaborate with AI while keeping engineering judgment intact".
The market gap is clear: no off-the-shelf product offers a unified experience covering a real multi-file IDE, multiple AI models, real-time interviewer monitoring, AI-interaction auditing, and structured Checkpoint grading. Meta had to heavily customize CoderPad — most companies don't have those engineering resources.
This project closes that gap. It's a self-hostable platform that mirrors Meta's interview format end-to-end, so any team can evaluate candidates the way modern engineers actually work.
HR / interviewers manage interviews, questions, and the platform from a single back-office console.
Pick a question, schedule a slot, and generate a one-click invitation link in seconds.
Each interview produces a unique, expiring invitation URL for the candidate.
Candidates land on a clean invitation page that explains the rules before they join.
A full multi-file IDE with file tree, CodeMirror 6 editor, test-result panel, integrated terminal, and a side AI assistant with live model selection (Gemini, GPT, etc.).
- Multi-file project IDE built on CodeMirror 6 with file tree and tabs — not a single-file LeetCode box.
- Governed AI assistant with streaming responses (AI SDK v6 + Spring AI). The interviewer chooses which model is allowed per session.
- Checkpoint-driven progression — each question is split into stages (debug → implement → optimise) following Meta's three-phase format.
- Real test execution in a sandboxed Docker container, with live test output and an interactive terminal (xterm.js + WebSocket).
- Auto-advancing flow: pass a checkpoint → next stage unlocks → finish all → completion screen.
- One-click interview creation with question selection, scheduling and AI-model policy.
- Invitation system with unique tokens and configurable expiry.
- Interview lifecycle management:
SCHEDULED → ACTIVE → COMPLETED / CANCELLED, with state transitions enforced at the domain level. - Conversation history & checkpoint results are persisted for post-interview review.
- Pilot AI scoring (preview) — an LLM judge produces a structured agency score against Meta's evaluation rubric.
- Self-hostable: one Docker Compose file brings up backend, two frontends, PostgreSQL and a sandboxed Docker-in-Docker for code execution.
- Spring Modulith modular monolith — clean module boundaries (
interview / execution / invitation / question / ai / scoring) verified byModularityTests, leaving the door open to split into services later. - Liquibase migrations, Spring Data JDBC (no JPA magic), and Testcontainers + Cucumber BDD scenarios out of the box.
┌─────────────────────┐ ┌─────────────────────┐
│ Admin (Next.js) │ │ Candidate (Next.js) │
│ port 3000 │ │ port 3001 │
└──────────┬──────────┘ └──────────┬──────────┘
│ Route Handlers (proxy) │
└────────────┬───────────────┘
▼
┌──────────────────────────────────┐
│ Spring Boot 4.0 — Modulith │
│ ┌────────────────────────────┐ │
│ │ interview / invitation / │ │
│ │ question / execution / │ │
│ │ ai / scoring │ │
│ └────────────────────────────┘ │
└────┬──────────┬─────────┬────────┘
▼ ▼ ▼
PostgreSQL Docker Gemini /
17 + Liqui- sandbox Spring AI
base (DinD) ChatClient
| Layer | Choice |
|---|---|
| Backend | Spring Boot 4.0, Java 25, Spring Modulith, Spring Data JDBC |
| AI | Spring AI 2.0.0-M4 (Gemini 3 Flash Preview by default) |
| Database | PostgreSQL 17 + Liquibase |
| Code execution | Docker (docker-java); roadmap: Firecracker microVM |
| Frontend | Next.js 16.1 (App Router), TypeScript, Tailwind CSS 4 |
| Editor | CodeMirror 6 (@uiw/react-codemirror) |
| AI UI | AI SDK v6 (useChat + DefaultChatTransport) + AI Elements |
| Terminal | xterm.js + WebSocket |
| Monorepo | npm Workspaces (apps/admin, apps/candidate, packages/shared) |
| Testing | JUnit 5, Mockito, Testcontainers, Cucumber 7 BDD |
ai-coding-interview/
├── backend/ # Spring Boot 4 / Java 25 (Gradle Kotlin DSL)
│ └── src/main/java/com/interview/
│ ├── interview/ # Interview aggregate, checkpoint progress
│ ├── invitation/ # Invitation tokens & candidate join flow
│ ├── question/ # Questions loaded from classpath YAML
│ ├── execution/ # Sandboxed Docker code execution
│ ├── ai/ # Spring AI ChatClient + streaming SSE
│ └── scoring/ # LLM judge for agency scoring (preview)
├── frontend/ # npm Workspaces monorepo
│ ├── apps/admin/ # Interviewer app (port 3000)
│ ├── apps/candidate/ # Candidate app (port 3001)
│ └── packages/shared/ # Shared types, API client, UI elements
├── docs/ # PRD, design docs, research notes
├── images/ # README screenshots
├── backend/compose.yaml # PostgreSQL for local backend dev
├── docker-compose.yml # Local full-stack runtime
└── build-and-run.sh # One-shot: build images and start everything
Read the local developer setup guide for API keys, building exams/100, loading it into DinD, deployment and validation. Use JDK 25; frontend development uses Node 24/npm.
建置需要 JDK 25 與正在執行的 Docker；前端開發模式另外需要 Node 24/npm。
本機測試 repo 的考題時，先從根目錄編譯：
docker build -t spike19820318/ai-coding-interview-question01:latest exams/100接著依手冊第 3 章建置平台、將考題 image 載入 DinD，最後啟動服務。Host Docker 與 DinD 的 image cache 不共用。
./build-and-run.sh 可建置平台並下載遠端現成考題，但不會編譯 exams/100；要驗證本機考題修改，請走手冊主線。
| 入口 | 網址 |
|---|---|
| 面試官 | http://localhost:3000 |
| 候選人 | http://localhost:3001（使用面試邀請連結進入） |
| Backend | http://localhost:8080 |
cd backend
./gradlew test # everything (needs Docker)
./gradlew test --tests "com.interview.*.bdd.*" # Cucumber BDD scenarios
./gradlew test --tests "com.interview.ModularityTests" # module boundary checksImplemented today (MVP):
- ✅ Interview lifecycle, invitations, multi-file Checkpoints
- ✅ Sandboxed Docker code execution + interactive terminal
- ✅ Streaming AI chat with dynamic model selection
- ✅ Pilot LLM-as-judge agency scoring
Next up:
- 🔮 Real-time interviewer "follow the candidate" mode + AI-interaction live panel
- 🔮 Full session replay (keystroke-level, multi-track, variable speed)
- 🔮 Question authoring UI + AI-assisted difficulty calibration
- 🔮 Anti-cheat signals (tab-switch, external paste, behavioural analysis)
- 🔮 Firecracker microVM execution backend
See docs/PRD.md for the full product spec.
Issues and pull requests are very welcome. Before opening a PR:
- Run
./gradlew testand make sure the Modulith verification still passes. - New backend behaviour should come with a Cucumber feature in
backend/src/test/resources/features/. - Follow the coding principles in
CLAUDE.md: readability first, expressive names, comments explain why, stateless services.
Apache License 2.0 — see LICENSE.