ZenCV: AI-Powered Job Application in 3 Clicks

The Problem

Job hunting is brutal. You find a great posting, spend 45 minutes tweaking your CV to match the keywords, write a cover letter that doesn’t sound like a robot wrote it, and then do it all over again for the next one. Multiply that by 50 applications and you’ve lost a week of your life.

I wanted to fix that.

What is ZenCV?

ZenCV is a Chrome browser extension backed by a Python/FastAPI backend, that automates the most soul-crushing part of job hunting.

Open a job posting. Click the extension. Get a tailored CV and cover letter in seconds.

That’s the whole pitch. Three clicks.

How It Works

Job Posting Page
      |
[ZenCV Chrome Extension]  ← React + TypeScript
      |
      ↓
POST /api/job-description/process
      |
[ZenCV Server]  ← FastAPI (Python) or IntelliCV Server (Node.js)
      |
      ↓
LLM (gpt-4.1-mini)
      |
Tailored CV JSON + Cover Letter
      |
[Extension renders templates + export]

Extract: The extension grabs and cleans the job posting HTML from your active tab using Chrome’s scripting API.
Match: The backend scans your stored cv_variants.json (multiple versions of your experience) and picks the most relevant bullets, skills, and summaries.
Generate: An LLM rewrites and tailors the content to match the job description’s language and priorities.
Export: Choose from 6 CV templates, preview, and export via browser print or draft directly into Gmail.

The Repos

Repo	Stack	Role
`ZenCV-client`	React, TypeScript, Tailwind, Vite	Chrome extension UI
`ZenCV-server`	Python, FastAPI, OpenAI SDK	Primary AI backend
`intellicv-server`	Node.js	Alternative/original backend implementation

The client is backend-agnostic, just point VITE_API_BASE_URL at whichever server you’re running.

Tech Stack Deep Dive

ZenCV Client (Chrome Extension)

React + TypeScript for the popup UI
Tailwind CSS for styling across 6 CV templates
Vite for build tooling with hot-reload dev mode
Chrome’s scripting API to reach into the active tab’s DOM
Gmail integration to open cover letters as compose drafts

ZenCV Server (Python / FastAPI)

FastAPI for clean, auto-documented REST endpoints (/docs ships with Swagger UI)
OpenAI-compatible: works with any provider supporting the OpenAI API spec
Multi-variant CV system: store different versions of your professional history; the LLM picks the best fit
Artifact caching by job URL hash so repeated requests don’t cost tokens
Token usage + cost tracking in every response summary

IntelliCV Server (Node.js)

The original backend implementation
Same API contract, different runtime. Useful if you prefer staying in the JS ecosystem.

The Data Model

The key is cv_variants.json. Instead of one CV, you maintain multiple variants of each section:

{
  "variants": [
    {
      "label": "backend-focus",
      "summary": "Backend engineer with 4 years in distributed systems...",
      "skills": ["Python", "PostgreSQL", "Docker", "Kubernetes"],
      "experience": [...]
    },
    {
      "label": "fullstack-focus",
      "summary": "Fullstack developer comfortable across the entire stack...",
      "skills": ["React", "TypeScript", "Node.js", "FastAPI"],
      "experience": [...]
    }
  ]
}

The LLM reads the job description, picks the right variant, rewrites the bullets to echo the posting’s language, and outputs a structured CV JSON ready for the template engine.

What I Learned

LLM prompt engineering is 80% of the work. Getting the model to select the right variant, stay factually grounded (no hallucinated experience), and match the tone of the job posting without being sycophantic took a lot of iteration.

Chrome extensions are surprisingly painful. Manifest V3, content security policies, and the quirks around chrome:// pages make development feel like navigating a minefield. The real-time extension reloading setup in dev mode saved my sanity.

Two backends, one contract. Maintaining both a Python and Node.js server forced me to think hard about API design upfront. The REST contract is clean enough that swapping backends is a single env var change.

Cost matters. I added token tracking from day one. Every generated artifact writes a summary.json with usage stats. Even at gpt-4.1-mini prices, it adds up across hundreds of applications, and knowing the cost per generation shaped prompt length decisions.

What’s Next

PDF export without the print dialog hack
Authentication so the server can be hosted publicly
Smarter variant selection using embedding similarity rather than pure LLM reasoning
Support for non-English job postings