Catalyst: AI-Powered Skill Assessment & Personalised Learning Plan Agent

A resume tells you what someone claims to know — not how well they actually know it. Catalyst is an AI-powered agent built with LangGraph, LangChain, and Streamlit that takes a Job Description (JD) and a candidate's resume, conversationally assesses real proficiency on each required skill, identifies gaps, and generates a personalised learning plan focused on adjacent skills the candidate can realistically acquire — with curated resources and time estimates.

---

🚀 Working Prototype & Setup

Local Setup Instructions

1. Clone the repository:

   git clone <repository-url>
   cd catalyst-project-main

2. Create a virtual environment and install dependencies:

   python -m venv venv
   source venv/bin/activate  # On Windows use: .\venv\Scripts\activate
   pip install -r requirements.txt

3. Configure Environment Variables: Create a .env file in the root directory and add your Google Gemini API Key:

   GOOGLE_API_KEY=your_api_key_here
   

4. Run the Streamlit application:

   streamlit run app.py

Deployed URL: https://catalyst-project-sandeep-balabantaray.streamlit.app/

---

🎥 Demo Video

[Watch the 3-5 minute demo video walkthrough here](Insert YouTube/Loom Link Here)

---

🏗 Architecture Diagram

Catalyst leverages a multi-agent architecture orchestrated by LangGraph.

graph TD
    A[Upload Resume & JD] --> B(Skill Extractor Agent)
    B -->|Identify Gaps via Knowledge Graph & LLM| C(Interviewer Agent)
    C -->|Scenario-based Question| D[Human User]
    D -->|Answer| E(Evaluator Agent)
    E --> F{More Skills to Assess?}
    F -- Yes --> C
    F -- No --> G(Mentor Agent)
    G -->|RAG + Skill Graph + LLM| H[Output Personalised Learning Plan]

Loading

🧠 Scoring and Logic

Catalyst utilizes a sophisticated Hybrid Scoring Engine and Skill Knowledge Graph to ensure deterministic and objective assessments, preventing LLM hallucinations.

1. Skill Extraction (skill_extractor):

   • Deterministic Graph Matching: First queries a NetworkX Skill Knowledge Graph to extract skills required in the JD that are missing from the resume.
   • LLM Fallback: If the graph yields fewer than 3 gaps, it uses Gemini to intelligently extract remaining core technical skills.

2. Assessment Loop (interviewer & evaluator):

   • Interviewer: Generates open-ended, scenario-based interview questions to probe for actual hands-on experience rather than trivia definitions.
   • Evaluator (Hybrid Scoring Engine): Scores answers on a 0-5 scale and assigns proficiency levels (Novice to Expert) using a weighted 3-signal approach:
     • LLM Score (50%): Qualitative depth assessment by the LLM.
     • Semantic Similarity (30%): ChromaDB distance comparison against expected concepts.
     • Keyword Coverage (20%): Exact matching for expected technical terms.

3. Personalised Learning Plan (mentor):

   • RAG Resource Curator: Retrieves real, curated learning resources from a ChromaDB vector store. No hallucinated URLs.
   • Graph-based Advice: Suggests adjacent skills and prerequisites based on the Knowledge Graph.
   • Outputs a highly personalized 4-6 week markdown study plan with actionable weekly goals, curated resource links, and realistic time estimates.

---

Sample Inputs and Outputs

Sample Input (JD): "Looking for a Backend Python Developer experienced with Django, PostgreSQL, and REST APIs. Experience with Docker and CI/CD pipelines is a plus."

Sample Input (Resume): "Software Engineer with 2 years of experience. Built applications using Python and Flask. Some experience with SQL databases like MySQL."

Sample Output (Learning Plan Excerpt):

• Goal: Master Django ORM and REST Framework.
• Resource: Build a simple blog API using Django REST Framework (DRF) · Tutorial · ~15h
• Time Estimate: 15 hours.
• Goal: Transition from MySQL to PostgreSQL and containerize the app.
• Resource: Dockerize the Django blog API · Documentation · ~10h
• Time Estimate: 10 hours.