I built Nexaura AI Resume Intelligence Platform using RAG, LangChain, FAISS and FastAPI,Crew.AI

#ai #fastapi #python #machinelearning

What is Nexaura?

Nexaura is a full-stack AI Resume Intelligence Platform I built
completely from scratch. Upload your PDF resume and get instant
AI analysis, ATS scoring, career coaching, and interview preparation.

Live Demo: https://ai-resume-intelligence-platform-two.vercel.app

GitHub: https://github.com/Sasanksurya/AI-Resume-Intelligence-Platform

The Problem

75% of resumes are rejected by ATS systems before a human ever
sees them. Job seekers don't know which skills are missing and
get no personalized feedback. Nexaura solves this completely.

Features

Resume Upload — PDF text extraction using PyMuPDF
Resume Summary — AI extracts Name, Skills, Experience, Education
AI Resume Chat — Ask anything about your resume using RAG pipeline
Quick ATS Score — Match resume against any job description instantly
Deep Analysis — Custom 4-Agent AI Pipeline
Interview Question Generator — 10 tailored questions from your resume
Resume Improvement Suggestions — AI powered career advice

The 4-Agent AI Pipeline

The most powerful feature is the custom multi-agent system:

Agent 1 — Resume Analyzer
Reads your full resume and extracts all key information with high accuracy

Agent 2 — JD Matcher
Compares resume against job description and identifies matched and missing skills

Agent 3 — ATS Scorer
Calculates ATS score from 0 to 100 based on keyword density and skill matches

Agent 4 — Career Coach
Gives 5 specific actionable suggestions to improve your resume for the exact role

Each agent passes its output to the next agent as context making
the final result much more accurate than a single AI call.

Tech Stack

Frontend:

Next.js 14
TypeScript
Tailwind CSS
Deployed on Vercel

Backend:

FastAPI (Python)
LangChain
FAISS Vector Database
RAG Pipeline
Groq AI with llama-3.3-70b-versatile
TF-IDF Embeddings with Scikit-learn
PyMuPDF for PDF extraction
Deployed on Render

System Architecture

How it works step by step:

User uploads PDF resume
PyMuPDF extracts all text from PDF
LangChain splits text into smaller chunks
TF-IDF converts chunks into vector embeddings
FAISS stores vectors in memory and on disk
When user asks a question the RAG pipeline searches FAISS
Most relevant chunks are passed to Groq LLM as context
Groq generates accurate response based only on resume content
Response is displayed on the Next.js frontend

This RAG architecture ensures the AI never hallucinates —
it can only answer from your actual resume content.

Why I Used TF-IDF Instead of BERT

Render free tier has a 512MB memory limit. Heavy embedding
models like BERT or sentence-transformers require PyTorch
which is 800MB alone. I built a custom TF-IDF embedding class
that extends LangChain base Embeddings — extremely lightweight,
no GPU needed, runs reliably on free infrastructure.

This was an important engineering decision that kept the entire
system running within free tier limits.

Deployment

Frontend is deployed on Vercel with automatic deployments
from GitHub main branch.

Backend is deployed on Render as a Python web service.
Environment variables for GROQ_API_KEY are stored securely
in Render dashboard.

Both are production live right now — not a localhost demo.

API Endpoints

Method	Endpoint	Description
POST	/api/upload	Upload PDF resume
GET	/api/resume-summary	Get AI resume summary
POST	/api/chat	Ask question about resume
POST	/api/ats-score	Quick ATS analysis
POST	/api/ats-score-crew	Deep 4-agent ATS analysis
GET	/api/interview	Generate interview questions
GET	/api/improve	Get resume improvements