top of page

HR Portal with AI-Powered Chatbot

  • Photo du rédacteur: dov azogui
    dov azogui
  • 11 sept.
  • 3 min de lecture

Internship at Ernst & Young (EY) — January 2025 to May 2025Location: Tel Aviv, IsraelRole: AI & Software Engineer Intern

Project Context

As part of my internship in the Technology Consulting division at EY Israel, I was assigned to develop a full-scale internal HR Portal powered by artificial intelligence. The objective was to create a tool that would allow EY employees to query complex HR documentation using natural language, instead of browsing through dozens of PDFs or contacting HR staff directly.

The challenge was to build a solution that could understand context, handle vague or misspelled questions, and provide clear, grounded responses—ideally with links or references to the original HR documents.

Problem Solved

EY Israel's HR department serves close to 2,000 employees. These employees often had difficulty locating accurate and up-to-date information about policies, onboarding procedures, vacation rights, travel reimbursement, parental leave, etc.

The existing system was based on a SharePoint file repository with hundreds of documents in Hebrew and English. Navigating it was time-consuming and not intuitive.

The solution was to replace this search process with a smart AI chatbot embedded directly in the HR Portal.

Technical Implementation

1. Architecture Overview

The solution is built on a full-stack architecture:

  • Frontend: React.js, integrated into SharePoint via Power Apps and custom components.

  • Backend: Python (Flask), REST API architecture, with business logic and logging.

  • Database & Indexing: SQLAlchemy for structured data, Azure Cognitive Search for semantic indexing.

  • AI Layer: Azure OpenAI’s GPT models used with Retrieval-Augmented Generation (RAG) pipeline to generate grounded answers based on real HR content.

2. Semantic Search and RAG

At the core of the system is a semantic search pipeline powered by Azure Cognitive Search. Internal documents (PDFs, Word files, internal procedures) were indexed with vector embeddings.

When a user submits a question, the system:

  • Extracts the semantic meaning of the question using an embedding model.

  • Queries the index to retrieve the top 3–5 most relevant passages from the HR documents.

  • Sends these retrieved texts as context to the GPT model using a RAG (Retrieval-Augmented Generation) strategy.

  • Returns a natural-language answer, with the option to open the original document if needed.

This ensures answers are grounded in official EY content and not hallucinated.

3. Production Deployment

The solution was built with deployment and scalability in mind:

  • Deployed using Docker containers in a production Azure environment.

  • CI/CD pipelines (GitHub Actions) ensured automated testing and deployment on merge to main.

  • Logging and monitoring systems were put in place to track usage patterns, API response times, and failure rates.

The chatbot was embedded in EY’s internal SharePoint portal using an iframe and exposed via Power Apps, making it accessible to all employees without any technical barriers.

Results and Impact

  • Reduced HR workload: Employees no longer need to contact HR for basic policy questions.

  • Improved employee satisfaction: Response time dropped from hours or days to seconds.

  • Scalable and maintainable: Built in a way that HR admins can upload new documents, which are automatically re-indexed and integrated into the chatbot.

  • Language flexibility: The solution supports queries in both Hebrew and English, handling mixed-language prompts as well.

  • Reliable grounding: All answers include a link to the source document used to generate the response, ensuring full transparency and trust.

Key Contributions

  • Designed and implemented the end-to-end AI chatbot system from scratch, including backend services, integration logic, and user interface.

  • Built and tuned the semantic search pipeline, carefully balancing relevance and speed using custom ranking profiles in Azure Cognitive Search.

  • Developed API communication between the React frontend and Flask backend, with proper error handling and fallback logic.

  • Managed all infrastructure deployments with Docker, ensuring smooth delivery of new features via CI/CD.

  • Worked directly with HR stakeholders to identify user requirements, perform user testing, and gather feedback for continuous improvement.

Final Thoughts

This project was a hands-on opportunity to build an AI-powered product with real business impact. It taught me how to design a full-stack solution that is user-focused, robust, and scalable—while integrating advanced AI capabilities responsibly. The project is currently live and used daily by the entire EY Israel team, demonstrating how applied AI can solve real-world inefficiencies in large organizations.

 
 
bottom of page