Building LinkedIn’s GenAI Platform — Xiaofeng Wang



AI Summary

Summary of Building LinkedIn’s JI Platform

Introduction

  • Presentation by Sh, Manager of JI Foundation.
  • Overview of the journey in developing the JI platform, including its significance.

Journey of the JI Platform

  1. Initial Launch (2023)
    • Launched collaborative articles using GPT-4 for content generation.
    • Developed gateway for centralized model access and prompt engineering tools.
    • Utilized Java for online phase and Python for backend.
  2. Second Generation Development
    • Introduction of “Co-pilot” feature to provide job fit recommendations.
    • Transitioned to a unified tech stack using Python.
    • Invested in prompt management to create a source of truth for developers.
    • Developed conversational memory to track interactions and enhance user experience.
  3. Launch of LinkedIn Assistant
    • Created a multi-agent system to automate tasks for recruiters, such as job postings and candidate evaluations.
    • Enhanced Python SDK for distributed agent orchestration, enabling more complex interactions.
    • Implemented a skill registry for easy API publishing and invocation.
  4. Memory Management Enhancements
    • Extended capabilities to experiential memory for knowledge extraction and analysis.
    • Organized various memory types for improved agent context awareness.
  5. Operational Insights
    • Developed in-house solutions for tracking and analyzing operational data for better agent performance.

Key Components of the JI Platform

  • Four Layers: Orchestration, Prompt Engineering Tools, Skills Invocation, Memory Management.
  • Ensured best practices and governance via a centralized platform to facilitate developer efficiency and responsibility.

Importance of Building the Platform

  • Bridged gaps between AI engineers and product engineers, emphasizing the need for a unified approach in the evolving AI ecosystem.
  • Stressed the uniqueness of JI systems over traditional AI systems, where optimization and serving phases merge.

Team Building Recommendations

  • Hire strong software engineers with integration skills, prioritizing potential over experience.
  • Build diverse teams to promote skill development through collaboration.
  • Emphasize critical thinking due to the fast-paced nature of the field.

Key Takeaways

  1. Tech Stack: Recommended use of Python for better scalability and community support.
  2. Essential Components: Develop a robust prompt version control and memory system.
  3. Adoption Strategies: Start with minimal viable products, scale based on immediate needs, and focus on user experience for developer productivity.

Additional Resources

  • For more technical details, refer to LinkedIn’s engineering blog posts by Sh and team.