INTERVIEW COPILOT — AI INTERVIEW PREPARATION

THE PROBLEM

Most interview preparation is generic. Candidates study broad question banks while the interview itself is shaped by the specific gap between their resume and that particular job description. The useful question is not "what are common interview questions" but "given this resume and this role, what will they actually ask me, and what am I missing".

Interview Copilot turns a resume and a job posting into a concrete, personalised preparation plan.

WHAT IT DOES

The candidate uploads a resume and pastes a job description. The application returns a tailored preparation report containing:

- a match score between the resume and the role

- a skill-gap analysis showing what the posting expects that the resume does not evidence

- likely interview questions with model answers

- a structured study plan

- a downloadable PDF of the full report

HOW IT WORKS

1. Ingestion. The resume is uploaded through Multer and parsed with pdf-parse to extract its text.

2. Analysis. The extracted resume text and the job description are sent to Google Gemini (gemini-2.5-flash) for comparison and report generation.

3. Validation. The model response is constrained to a strict JSON shape and validated before it is trusted.

4. Rendering. The React front end renders the report, and Puppeteer generates a downloadable PDF version server-side.

ENGINEERING DECISIONS

- Structured model output. Free-form LLM text is unreliable to build a UI on, so the Gemini call is constrained to a fixed JSON schema using responseSchema, and the response is validated again with Zod before it reaches the front end. If the model returns something unexpected, it is caught at the boundary rather than breaking the interface.

- Authentication. Sessions use JWTs delivered in HTTP-only cookies, which keeps the token out of reach of client-side JavaScript. Passwords are hashed with bcryptjs. Logged-out tokens are placed in a MongoDB blacklist collection with a TTL index, so expired entries are removed automatically rather than accumulating.

- Separation of concerns. The React single-page application and the Express API are deployed independently, on Vercel and Render respectively, communicating over a REST API.

TECH STACK

Frontend: React 19, React Router 7, Vite, Context API, Sass, Axios

Backend: Node.js, Express 5, MongoDB Atlas, Mongoose, JWT, bcryptjs, Multer, pdf-parse

AI and documents: Google Gemini API, Zod, Puppeteer

Deployment: Vercel (frontend), Render (backend)

STATUS AND LINKS

Deployed and publicly accessible.

Live: https://interview-copilot-ai-iota.vercel.app/

Source: https://github.com/gativarshney/interview-ai