iTranslated by AI

The content below is an AI-generated translation. This is an experimental feature, and may contain errors. View original article
🌤️

AI Mock Interview App Released

に公開
2025/09/12
1件

1. Introduction

I have released "Skymock AI," a mock interview application for IT engineers.
No sign-in is required, so anyone can use it easily.
https://skymockai.com/

Skymock AI is a service that allows for role-specific real-time mock interviews.
In this article, I would like to introduce the functional overview and technical configuration of Skymock AI.

2. Functional Overview

Skymock AI currently provides the following features:

  • Real-time Mock Interviews: Mock interviews are possible through real-time voice interaction. It currently supports interviews for the following roles:

    • Front-end Engineer
    • Back-end Engineer
    • Full-stack Engineer
    • SRE
    • Native App Engineer
    • Machine Learning Engineer

  • Interview History Storage: All mock interview contents are saved, allowing users to look back at them anytime.

  • Interview Feedback: After the interview, feedback is provided to the user based on the interview content. You can review specific feedback regarding strengths and areas for improvement.

  • Multilingual Support: Currently supports English and Japanese.

3. Technical Configuration

Front-end

For the front-end, I have adopted Next.js App Router.

  • useSWR: A data fetching library using React Hooks that allows for efficient data retrieval from APIs and cache management.
  • React Hook Form: A library for efficiently managing form state and simplifying validation. It allows for simple implementation of form state management, input validation, and submission processing.
  • Zod: A type-safe schema validation library used for validating form inputs and API response data.
  • shadcn/ui: A UI library based on Radix UI (Headless UI) and Tailwind CSS. shadcn/ui is not added to npm dependencies; instead, components are copied locally. This makes it easy to extend the UI.
  • Framer Motion: A library that allows for easily adding animation effects to the UI.
  • Babylon.js: A library for achieving advanced 3D graphics, used for displaying the interviewer's avatar. Since I hadn't touched WebGL-based libraries before, I decided to give it a try.

Back-end

The back-end adopts FastAPI, a Python web framework.
The main libraries used are as follows:

  • Pydantic: A library for data validation and settings management in Python. It has high affinity with FastAPI and is utilized for defining data models for API requests and responses, validation, and serialization.
  • SQLAlchemy: An ORM (Object Relational Mapper) library for Python.
  • Alembic: A database migration tool that works with SQLAlchemy. It manages database schema changes and, like a version control system, can record, apply, and roll back schema change history.

LLM Infrastructure (Groq)

LLMs are used to generate interview questions and provide feedback. I use Groq for the LLM API.
Groq is a company founded by Jonathan Ross, a former Google engineer. It provides high-speed, low-cost LLM services using a proprietary semiconductor chip called the LPU (Language Processing Unit), designed to accelerate inference for Large Language Models (LLMs).
https://groq.com/

Groq offers a variety of models, allowing you to select the most suitable one for your use case. Furthermore, the API usage fees are reasonable, enabling the use of high-quality LLMs while keeping costs down.
https://console.groq.com/docs/models

Audio Implementation

Skymock AI's voice bot is implemented by combining an LLM with TTS (Text-to-Speech) and STT (Speech-to-Text) instead of using the Realtime API to keep costs down.

TTS (Text-to-Speech)

I have adopted OpenAI's TTS model for generating the AI interviewer's voice. High-quality speech synthesis and natural intonation provide users with a realistic interview experience. Multiple voice styles are available, allowing for the application of the most suitable voice based on the interviewer's character setting.

STT (Speech-to-Text)

I have adopted Groq's Whisper API for user voice recognition. Groq's Whisper API provides lower-cost and lower-latency voice recognition compared to OpenAI's Whisper, offering an excellent balance between real-time performance and cost-effectiveness.
The following benchmarking was very helpful for comparing STT API performance:
https://artificialanalysis.ai/speech-to-text/models/whisper

Voice Recognition on the Browser

I use an OSS called vad to detect the start and end of speech on the browser.
https://github.com/ricky0123/vad

vad is a JavaScript library that uses VAD (Voice Activity Detection) technology to detect speech segments from audio signals. It allows for the implementation of audio-related processing, such as displaying the microphone permission popup, starting recording, and detecting speech status, with a simple interface.

For example, voice activity can be detected simply by calling the useMicVAD hook as follows:

const vad = useMicVAD({
  startOnLoad: true,
  onSpeechEnd: (audio) => {
    const wav = utils.encodeWAV(audio);
    const blob = new Blob([wav], { type: "audio/wav" });
    send(blob);
  },
  positiveSpeechThreshold: 0.6,
  workletURL: "/vad/vad.worklet.bundle.min.js",
  modelURL: "/vad/silero_vad.onnx",
  ortConfig(ort) {
    ort.env.wasm = {
      wasmPaths: {
        "ort-wasm-simd-threaded.wasm": "/vad/ort-wasm-simd-threaded.wasm",
        "ort-wasm-simd.wasm": "/vad/ort-wasm-simd.wasm",
        "ort-wasm.wasm": "/vad/ort-wasm.wasm",
        "ort-wasm-threaded.wasm": "/vad/ort-wasm-threaded.wasm",
      },
    };
  },
});

Authentication Implementation

Implemented by combining Auth.js and Firebase Auth.

Database

For the database, I adopted Neon, a Postgres-compatible serverless DB.
I was torn between Neon and Supabase, but I chose Neon because its free plan allows for branching features.
Neon's paid plan is also $6/month cheaper than Supabase's.

https://neon.tech/

Hosting

For deployment, I have adopted Google Cloud's Cloud Run for both the back-end and front-end.
At Cloudflare's Builder Day 2024, it was announced that Next.js could now run on the Node runtime, so I wanted to use Cloudflare. However, an error occurred during the Firebase Admin SDK build, so I used the "stable?" choice, Cloud Run.

CI/CD

The CI/CD pipeline uses GitHub Actions and Cloud Build.
GitHub Actions executes automated tasks such as builds, tests, and deployments triggered by code changes. On the other hand, Cloud Build handles the building, testing, and deployment of container images.
Google Secret Manager is used for secret management.

The deployment flow is as follows:

  1. PR Merge: When a Pull Request is merged from a development branch into the main branch, GitHub Actions is triggered, and the Release Please workflow is executed. Release Please automatically generates a release Pull Request based on semantic versioning. This PR includes version updates, CHANGELOG updates, and more.
  2. Release PR Merge: When the release PR is merged into the main branch, the Cloud Build trigger is activated.
  3. Build and Deploy: In Cloud Build, the latest code is built based on the Dockerfile to generate a container image. During the build process, authentication information is retrieved from Secret Manager and set as environment variables. The generated container image is then deployed to Cloud Run.

Other

  • Favicon Settings:
    I used RealFaviconGenerator to set up the website icons. It allows for the easy generation of favicons compatible with various devices and browsers.

https://realfavicongenerator.net/

  • Creating Release PRs:
    I use Release Please to create release PRs. Release Please automatically generates a Pull Request for release based on semantic versioning. This PR includes version number updates, CHANGELOG updates, and more.
    It's quite handy, so I personally find it very useful.

https://github.com/googleapis/release-please

5. Conclusion

In this article, I introduced "Skymock AI," a mock interview application for IT engineers.
Since only the minimum necessary features are currently implemented, I hope to expand the features and roles in the future.
If you have any feedback or opinions, I would appreciate it if you could contact me via DM at the following address :bow:

https://x.com/_takigon

Discussion

ちだちだ

こちらのアプリケーションについてご相談がありXの方フォローさせていただきました。
DMが閉鎖されているようですのであけていただけると幸いです。
https://x.com/Rikuto_Kaihatsu