Skip to content
BranchesTags

Latest commit

 

History

11 Commits

Folders and files

NameName
Last commit message
Last commit date
backend
backend
 
 
frontend
frontend
 
 
README.md
README.md
 
 

Repository files navigation

Mentor-AI

Chat with AI versions of Hitesh Choudhary and Piyush Garg — two Indian tech educators — and get answers that sound like them, not a generic chatbot.

Image

Live Demo GitHub Next.js Express OpenAI MongoDB TypeScript

What is this?

Mentor-AI is a persona-based chat application built for the GenAI with JS 2026 assignment. Instead of one generic assistant, you pick a mentor — Hitesh or Piyush — and every reply is shaped by that person's real speaking style, background, and teaching approach.

Sign in with Google, choose a persona, start a conversation, and ask anything about coding, GenAI, careers, or building projects. The app remembers context within a chat and routes each message to the correct persona prompt on the backend.

Link
Live app https://mentor-ai-ruddy.vercel.app
API https://mentor-ai-l4nc.onrender.com
Repository https://github.com/KumarNirupam1/Mentor-AI
Author Kumar Nirupam

Features

  • Two distinct AI personas — Hitesh Choudhary and Piyush Garg with separate system prompts
  • Persona switching — pick your mentor from the dashboard; each chat is tied to one persona
  • Google OAuth — secure sign-in with JWT stored in httpOnly cookies
  • Persistent chat history — create, list, resume, and delete conversations per persona
  • Context-aware replies — last 100 messages sent to the LLM for coherent multi-turn chat
  • Clean chat UI — landing page, login, persona picker, chat list, and live chat interface
  • Per-mentor rate limits — 10 messages per mentor, then a 10-minute cooldown (see below)
  • Bring your own OpenAI key — each user adds their own API key to chat; no shared server key

Rate limiting

To keep OpenAI costs under control, each logged-in user gets 10 messages per mentor (Hitesh and Piyush tracked separately). After the 10th message, chat is paused for 10 minutes with a countdown in the UI. Limits reset automatically when the cooldown ends.

Setting Default
Messages per mentor 10
Cooldown 10 minutes

Configurable on the backend via PERSONA_MESSAGE_LIMIT and PERSONA_COOLDOWN_MINUTES. Quota is reserved before each OpenAI call; failed replies do not count against the limit.

OpenAI API key (BYOK)

Each user must add their own OpenAI API key before chatting. Keys are encrypted at rest in MongoDB using AES-256-GCM (ENCRYPTION_SECRET on the server). They are never returned to the client. Hashing is not used — the server must decrypt the key to call OpenAI.

Architecture

User Browser (Next.js on Vercel)
        |
        |  Google OAuth / REST API (cookies)
        v
Express Backend (Render)
        |
        +-- MongoDB Atlas  (users, chats, messages)
        +-- OpenAI API     (gpt-4.1-mini, persona prompts)

Message flow

  1. User sends a message from the chat UI
  2. Frontend calls POST /api/v1/messages/:chatId with credentials
  3. Backend verifies auth, checks rate limit, loads chat, fetches last 100 messages
  4. Routes to hitesh() or piyush() in backend/src/utils/ai.ts
  5. Prompt = system persona + developer context + history + new message
  6. OpenAI reply saved to MongoDB and returned to frontend

Tech stack

Layer Technology
Frontend Next.js 16, React 19, TypeScript, Tailwind CSS 4
Backend Express 5, TypeScript, Mongoose
Database MongoDB Atlas
Auth Google OAuth 2.0, JWT (access + refresh tokens)
AI OpenAI gpt-4.1-mini
Deployment Vercel (frontend) + Render (backend)

Persona data — collection and preparation

Persona profiles were built from publicly available information about Hitesh Choudhary and Piyush Garg.

Sources

Source What we extracted
YouTube live streams and videos Speaking patterns, Hinglish mix, teaching style
Public social profiles Career history, companies, channel links
Known public quotes e.g. "Database dusre continent me hai"
Course and cohort announcements Topics taught, prerequisites explained
Assignment reference repos Base persona structure

Each persona prompt includes

  1. Identity and career background
  2. Personal traits, hobbies, favorite languages
  3. Public social and work links
  4. Speaking rules (Hindi vs English, greetings, tone)
  5. Real chat examples from live streams to anchor voice

No fine-tuning — personality comes from the system prompt + style examples.

Prompt engineering strategy

Layered system prompt

Role definition -> Background -> Traits -> Links -> Speaking rules -> Examples

Real style examples in the prompt

Hitesh:

  • "Kuch b use kro bs keys frontend me mt chipkana, otherwise you are on your own"
  • "Hanji! Swagat apka humare channel Chaicode par"
  • Long live-stream answers on MERN vs AI, courses, etc.

Piyush:

  • "Let's make our own llm", "Let's make our own cohort"
  • Hinglish live stream style, GenAI course pitches

Persona behavior rules

Hitesh Piyush
Greeting "Hanji!", Chai aur Code vibe "Hello hello", chai pe charcha
Language Hindi + English for tech Hinglish, "modren" not modern
Tone Humble experienced teacher Frank "make it from scratch" builder

Developer context

Each request adds: User which is asking the question is {name} and his email is {email}

Model

OpenAI gpt-4.1-mini — fast, follows long system prompts well.

Context management

Decision Implementation
History window Last 100 messages per chat
Storage MongoDB Chat + Message collections
Persona lock Each chat stores persona: hitesh or piyush
Chat title Auto-generated from first 3 words of first message
Auth JWT in httpOnly cookies with silent refresh

Sample conversations

Hitesh — MERN vs GenAI

You:    MERN pehle seekhun ya GenAI?

Hitesh: Dekho, software development skill hai jo aapko seekhni padegi.
        Interface to lagega na — chahe AI seekho ya ML, authentication to
        lagega hi. Bina HTML, JS ke project kaise banaoge?

Piyush — build from scratch

You:    Can we build our own LLM?

Piyush: Let's make our own llm — that's the whole vibe. Bas JavaScript
        aani chahiye, ek simple mon stack project ho, chahe todo app hi
        ho, even that is enough.

Project structure

Mentor-AI/
├── frontend/
│   ├── src/app/           Pages (landing, login, dashboard, chat)
│   ├── src/components/    ChatInterface, PersonaSelection, UI
│   ├── src/lib/           API helpers
│   └── public/            hitesh.webp, piyush.webp
├── backend/
│   └── src/
│       ├── controllers/   auth, chat, messages
│       ├── models/        User, Chat, Message
│       ├── routes/        REST API
│       └── utils/ai.ts    Persona prompts + OpenAI
└── README.md

Getting started

Prerequisites

  • Node.js 18+
  • MongoDB Atlas cluster
  • Google OAuth credentials
  • OpenAI API key

Run locally

git clone https://github.com/KumarNirupam1/Mentor-AI.git
cd Mentor-AI

# Backend
cd backend
cp .env.example .env
npm install
npm run dev

# Frontend (new terminal)
cd frontend
cp .env.example .env
npm install
npm run dev

Open http://localhost:3000, sign in with Google, pick a mentor, start chatting.

Environment variables

Backend (backend/.env)

NODE_ENV=development
PORT=8000
CORS_ORIGIN=http://localhost:3000
FRONTEND_URL=http://localhost:3000
MONGODB_URI=your_mongodb_connection_string
GOOGLE_CLIENT_ID=your_google_client_id
GOOGLE_CLIENT_SECRET=your_google_client_secret
GOOGLE_CALLBACK_URI=http://localhost:8000/api/v1/auth/google/callback
ACCESS_TOKEN_SECRET=your_random_secret
REFRESH_TOKEN_SECRET=your_random_secret
ACCESS_TOKEN_EXPIRY=1d
REFRESH_TOKEN_EXPIRY=15m

# No shared OpenAI key — users add their own via the app UI
ENCRYPTION_SECRET=your_long_random_encryption_secret_at_least_32_chars

# Optional rate limits (defaults: 10 messages, 10 min cooldown per mentor)
PERSONA_MESSAGE_LIMIT=10
PERSONA_COOLDOWN_MINUTES=10

Frontend (frontend/.env)

NEXT_PUBLIC_API_URL=http://localhost:8000

Never commit .env files.

Production

Service URL
Frontend (Vercel) https://mentor-ai-ruddy.vercel.app
Backend (Render) https://mentor-ai-l4nc.onrender.com

Production env:

GOOGLE_CALLBACK_URI=https://mentor-ai-l4nc.onrender.com/api/v1/auth/google/callback
FRONTEND_URL=https://mentor-ai-ruddy.vercel.app
CORS_ORIGIN=https://mentor-ai-ruddy.vercel.app
NEXT_PUBLIC_API_URL=https://mentor-ai-l4nc.onrender.com

Google Console: add origin https://mentor-ai-ruddy.vercel.app and redirect URI above.

API overview

Method Endpoint Description
GET /api/v1/healthcheck Server health
GET /api/v1/auth/google Start Google OAuth
GET /api/v1/auth/getme Current user
GET /api/v1/auth/logout Sign out
POST /api/v1/chats Create chat
GET /api/v1/chats?persona=hitesh List chats
DELETE /api/v1/chats/:id Delete chat
GET /api/v1/messages/:chatId Get messages
POST /api/v1/messages/:chatId Send message, get AI reply

Submission checklist

Requirement Link / Status
Live deployed website https://mentor-ai-ruddy.vercel.app
Public GitHub repo https://github.com/KumarNirupam1/Mentor-AI
LLM chat (both personas) Hitesh + Piyush via OpenAI
Persona switching Dashboard picker
Persona data docs See Persona data section
Prompt engineering See Prompt engineering section
Context management See Context management section
Sample conversations See Sample conversations section
Setup instructions See Getting started section

Evaluation alignment

Parameter Weight How addressed
Persona Accuracy 30 Real transcript examples, Hinglish rules, persona greetings
Conversation Quality 25 100-msg history, developer context, persona-locked chats
Technical Implementation 25 Clean REST API, separated ai.ts module, typed frontend
User Experience 20 OAuth flow, persona cards, chat list, live chat UI

Author

Kumar Nirupam — GitHub

Built for GenAI with JS 2026 — ChaiCode assignment.

About

Mentor-AI is a persona-based chat application built for the GenAI with JS 2026 assignment. Instead of one generic assistant

mentor-ai-ruddy.vercel.app

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages