CareerPath AI

An AI career counselor I built in 3 days as an experiment in coding with Gemini 3 Pro. It suggests careers that fit your profile, maps out how to get into each one, and generates 'day in the life' images of the job.

Description

CareerPath AI was my test of how far 'vibe coding' with Gemini 3 Pro could go. I built it over a single three-day weekend, and it works as a career counselor that tries to go past generic advice. You give it some background about yourself and take a quiz, and it suggests careers that fit you, each with a roadmap for getting there. It can also generate a realistic 'day in the life' slideshow for a career, so a job title turns into a picture of what the work looks like from day to day. It's a full-stack app with a React front end and a Python backend, and the AI wrote nearly all of it.

Media

1 of 9

The login page.

Features

Profile Analysis

Google Gemini 2.5 Flash reads through your profile and quiz answers and matches you with careers based on your aptitude and personality instead of keywords.

Career Roadmaps

Each career comes with a step-by-step plan for the education and experience you need to get there from where you are now, including advice for where you live.

Day-in-the-Life Slideshows

Google's Nano Banana image model generates a realistic slideshow of an ordinary day in a job, without glamorizing it, so you can get a feel for whether you'd actually want to do it.

Backend Proxy

The React front end never talks to Gemini directly. A Python FastAPI backend sits in between as a proxy and holds the API keys and the business logic.

Daily Limits

A quota system in the database caps how many times a day each user can run the expensive parts, like image generation and detailed career analysis.

Key Responsibilities

  • Built the full-stack codebase by prompting Gemini 3 Pro. I made the architecture decisions and the AI wrote the implementation.
  • Deployed the backend to Hugging Face Spaces with Docker and the front end to Vercel.
  • Designed the Supabase schema and Row Level Security policies so each user can only see their own data.
  • Set up image generation as parallel async API calls, which cut down how long users wait for their slideshows.
  • Used Zustand to manage state across the whole flow, from onboarding through to the results.

Additional Information

  • I built this to see how well Gemini 3 Pro could reason through a real project. It took 3 days, and the AI wrote nearly all of the code.
  • Most of what makes the app work is strict prompting that forces the LLM to answer in JSON, so the front end always gets structured data it can render.
  • The Gemini API keys stay on the backend proxy, so they never show up in the client-side code.
  • To save on tokens, career details load lazily. The detailed roadmap for a career is only generated when someone clicks into it.
  • The loading screens have a "Nano Banana" badge as a nod to the image model's codename.