Building an LLM-Powered Resume Screener Using Mistral and Gradio (Fully Offline)

Introduction

In today’s hiring landscape, HR teams often rely on Applicant Tracking Systems (ATS) to quickly filter resumes against job descriptions. While these systems work, they often fall short in understanding nuanced skillsets, context, and potential. Enter Large Language Models (LLMs). In this blog, I’ll walk you through how I built a smart resume screening app that mimics ATS behavior but leverages the power of an open-source LLM (Mistral), all running locally on your machine — no OpenAI keys, no API costs, no data leakage.

🔧 Tech Stack

• LLM: Mistral (7B model via Ollama)

• Interface: Gradio

• PDF Parsing: PyMuPDF (fitz)

• Language: Python

• Execution: Fully offline

---

📋 Prerequisites

• Python 3.8+

• Basic knowledge of how LLMs and APIs work (no prior ML experience needed!)

• 4GB+ RAM recommended

• OS: Windows, Linux, or macOS

⚙️ Setting Up Mistral via Ollama

To run the Mistral model locally for LLM inference, follow these steps:

1. Install Ollama

   • Go to: https://ollama.com/download

   • Download and install for your OS (Windows, Mac, Linux)

2. Pull the Mistral model

    ollama pull mistral
   

3. Run the Mistral model locally

    ollama run mistral
   

   This will start a local server at http://localhost:11434 which your app can query.

Note: You can replace mistral with other models supported by Ollama like phi, gemma, etc.

---

🌐 Key Features

• Upload any resume in PDF format

• Extracts structured details using LLM: name, experience, skills, work history, education

• Compares resume with job description and gives:

  • Match Score (0–10)

  • Strengths

  • Weaknesses

  • Suggested Interview Questions

• Displays results in a beautiful, responsive Gradio interface

• Shows processing time (performance benchmark)

---

🔄 Project Flow

1. Resume Upload

   • The user uploads a PDF resume

   • fitz (PyMuPDF) extracts raw text from the PDF

2. Resume Parsing

   • A prompt is generated asking the LLM to extract JSON structured data from the resume

   • The prompt is fed into a locally running Mistral model via Ollama (localhost)

3. Job Description Matching

   • The extracted resume JSON and pasted job description are passed to a second LLM prompt

   • The model analyzes the match, assigns a score, and lists strengths, weaknesses, and interview questions

4. Result Display

   • Everything is visualized in the Gradio app with markdown formatting

   • A timer displays how long the LLM took to process everything

🧰 Sample Prompt to LLM (Parsing Resume)

5.  You are an expert resume parser. Extract the following fields:
    - Name
    - Total years of professional experience
    - Skills
    - Work History
    - Education
    Return in JSON format.
   

   🔹 Sample Prompt to LLM (Matching with JD)

    You are an expert HR analyst. Given the candidate's resume (structured JSON) and a job description, analyze:
    1. Fit Score (0–10)
    2. Strengths
    3. Weaknesses
    4. Interview Questions
   

   📈 Example Output

   Parsed Resume:

{
  "name": "James Bond",
  "experience_years": 1,
  "skills": ["Python", "SQL", "Machine Learning", "HTML", "CSS", "Flask"]
}

Evaluation:

• Fit Rating: 6/10

• Strengths: Strong programming base, fast learner, good visual tools

• Weaknesses: No production-level experience, limited exposure to CI/CD

• Interview Questions: Describe a project using SQL; How would you build a scalable web app?

---

🚀 Why This Project Matters

• Replaces expensive black-box resume filters

• Great use of local inference with LLMs

• Emphasizes data privacy

• Easy for recruiters or HR tech tools to adopt

---

🏆 Bonus: Gradio UI Snapshot

The Gradio interface includes:

• File upload

• Textbox for job description

• Real-time results (markdown + timer)

• Mobile responsive

---

🚀 What's Next?

• Add bulk resume upload (batch evaluation)

• Export results to CSV/JSON

• Add support for multiple LLMs via config

• Integrate with ATS APIs or job boards

---

🔗 GitHub

View the Full Code on GitHub

---

📖 Final Thoughts

This project was a deep dive into how powerful open-source LLMs can be when used smartly. With just Python, a local model, and Gradio, you can build ATS-level functionality that respects privacy, costs nothing, and gives users total control.

Feel free to fork the repo, suggest features, or deploy it in your own use case!