JobAgent

JobAgent is a hybrid AI-driven solution designed to simplify and streamline the job application process by combining semantic evaluation, document parsing, and web-based company insights. JobAgent intelligently matches resumes against job descriptions, extracts relevant company information, and even drafts personalised or general outreach emails—reducing the time and effort required for job seekers.

🔍 Overview

JobAgent automates three critical parts of the job-seeking process:

1. Company Research

   • Creates concise company reports by referencing reviews and extracting key information from a company’s about page.
   • Performs sentiment classification on individual reviews and summarises both positive and negative perspectives.

2. Resume & Job Description Matching

   • Extracts and aligns hard skills, soft skills, years of experience, and responsibilities between resumes and job descriptions.
   • Scores the semantic relationship between the two, providing structured evidence of fit.

3. Personalised Communication

   • If a contact person is identified in the job description, JobAgent composes a personalised email.
   • Otherwise, it drafts a general application email using extracted context.

💡 Motivation

Searching for relevant jobs to apply for can be time-consuming and repetitive. Typically, candidates must:

• Review required qualifications and match them against their resume.
• Read role descriptions to check relevance.
• Explore company reviews and values to decide if it’s a good fit.
• Search for a contact person to send a tailored application.

JobAgent saves time and effort by automating these steps end-to-end, letting candidates focus on opportunities instead of logistics.

🔑 Key Concepts

1. Web Search & Company Insights

• Reviews & Overviews: Extracted asynchronously using TavilySearchResults.
• Sentiment Analysis: Individual reviews classified using siebert/sentiment-roberta-large-english.
• Summarization: Both company overviews and aggregated reviews (positive & negative) are summarised for quick insights.

2. Symantical Evaluation: Resume & Job Description

📄 PDFLoader

• Parses resumes (r) and job descriptions (job_d) into structured sections for direct comparison.
• Example sections:
  • Resume: Education, Skills, Professional Experience, Certificates, Volunteering, Projects
  • Job Description: Requirements, Responsibilities, Role Description, Contact Person

LangExtract

• A Google tool optimised for long-document entity extraction.
• Extracts hard skills, soft skills, years of experience, and contact persons with high recall by using chunking, parallel processing, and multi-pass strategies.

3. Hybrid Evaluation (Sentence Transformers + LLM)

JobAgent combines section-level comparisons and semantic embeddings:

A. PDFLoader-Based Evaluation

• Responsibilities ↔ Professional Experience (G-Eval):

  • Uses an LLM to score how well candidate experience matches stated responsibilities.

  • Scoring scale:

    • 0–0.4: Negative → clear mismatch.
    • 0.4–0.6: Neutral → inconclusive.
    • 0.6–1.0: Positive → strong match.

  • Job Title Matching

  • Embedding-based cosine similarity between job description titles and user-supplied role lists.

  • Threshold (alpha) ensures candidate searches align with role titles.

B. LangExtract-Based Evaluation

• Hard Skills:

  • Score = intersection of candidate’s hard skills with job requirements ➗ total of job requirements skills.
  • Predefined skills examples are provided here src/data/skills/skills.json.

• Soft Skills: /Users/galbeeri/Desktop/Private/Job Applications/JobAgent/job_agent/src

  • Uses predefined ontology (src/data/softskills/soft_skills.json) for mapping.
  • Evaluated via semantic similarity using TechWolf/JobBERT-v2, trained on 5M+ job-pairs.
  • Weighted dot product generates final soft_skill_score.

• Years of Experience:

  • Evaluated against job requirements.
  • Acceptable margin gap: <= 2 years.

• Contact Person:

  • If found in job description, email is personalised.
  • If not, a general but context-aware email is composed.

📧 Email Composition

Email drafts are created using a combination of:

• Overall evaluation scores (LangExtract + PDFLoader)
• Extracted company context (reviews + about-page summaries)
• Candidate-job alignment metrics

This ensures communication is relevant, personalised, and professionally tailored.

𐄷 Evaluation Notebooks

#	Name	Technique	View
1	search_sota_approach.ipynb	Sentiment Analysis, Text-Summarisation, and Agent Building
1	job_resume_match.ipynb	Semantic Evaluation, NER-LangExtract, and Agent Building

🧰 Tech Stack

• Document Parsing: PDFLoader-LangChain
• Entity Extraction: LangExtract
• Semantic Matching: Sentence Transformers (TechWolf/JobBERT-v2)
• LLM Evaluation: Contextual scoring of responsibilities and experience
• Sentiment Analysis: siebert/sentiment-roberta-large-english
• Web Search: TavilySearchResults
• Job Crawling: FireCrawl

🪾 Project Structure

.
├── config
├── LICENSE
├── notebooks
│   ├── job_resume_match.ipynb
│   └── search_sota_approach.ipynb
├── src
│   ├── agent
│   │   ├── __init__.py
│   │   └── state.py
│   ├── data
│   │   └── langextract
│   │       ├── skills
│   │       │   ├── skills.html
│   │       │   ├── skills.json
│   │       │   └── skills.jsonl
│   │       └── softskills
│   │           ├── soft_skills.json
│   │           ├── softskills.html
│   │           └── softskills.jsonl
│   ├── prompt_engineering
│   │   ├── __init__.py
│   │   └── prompts.py
│   └── utils
│       ├── __init__.py
│       ├── doc_utils.py
│       ├── langextract_utils.py
│       ├── search_utils.py
│       └── sentiment_utils.py
├── LICENSE
├── README.md
├── pyproject.toml
└── .gitignore