Ming Management Training, a personalized AI-native L&D coach

Transforming today’s technical talent into tomorrow’s executives using personalized AI-native training simulations and advanced coaching analytics.

Comment 1

What it does

Ming Management Training is an AI-native application that uses a multi-agent LLM architecture to help current and aspiring engineers-turned-managers practice managing real-world workplace scenarios. Users engage in role-playing difficult conversations (performance reviews, conflict resolution, or layoffs) with LLM-driven NPCs, receiving detailed analysis, actionable feedback, and new personalized scenarios based on their managerial strengths and weaknesses analyzed from the conversations.

What Inspired Us:

  • A desire to help people learn and build skills
  • As AI automation reshapes work, which jobs will endure, and how do we bridge displaced workers into roles the industry needs in this AI economy?
  • We realized that there is a lack of efficient and affordable training for the soft skills needed in technical managerial roles
  • High cost of hiring new managers compared to upskilling in-house technical employees, and may lack in-depth technical expertise, culture understanding and connection to the people in the company.
  • Desire to improve upon the generic and ineffective training solutions of the present (cough cough AlcoholEdu training)
  • What if engineers could be trained to become managers who combine deep technical expertise with excellent managerial skills?

What we learned:

  • How to implement scalable multi-agentic architecture
  • Effective teamwork makes the dream work! (fr tho)
  • The best approach to building something is just to get started - we had setbacks and dead ends, but we were always making progress
  • Supplemented human capabilities with LLM-targeted feedback
  • Plan ahead! 8 PM Friday is when you start coding, not when you start ideating

How we built our project:

  • First design server & multi-agentic architecture
  • Divide the task based on expertise and component structure (e.g., one for layout, another for conversation UI etc)
  • Next.js front-end with React, Tailwind, and ShadCN
  • CedarOS to integrate all of the above with all of the below
  • Mastra backend connected to Gemini APIs (LLM and Sentiment-Analysis)
  • Deployed full stack to Cloudflare, hosted on our own domain

The challenges we faced:

  • Shipping to Production: We initially bundled our Next.js and Mastra server as a single server with Cloudflare Wrangler, but had to split them into two separate servers and point our Next.js server at Mastra.
  • Implementing multi-agentic workflow & orchestrator LLM pipelines (see diagrams)
  • Validating our idea with an MVP - it was tricky knowing when we had actually achieved our MVP and deciding where to go next
  • Integrating variable outputs from LLMs caused us to rewrite our parser ~6 times
  • We tried live video generation, but it was too expensive (~$9/min) and we settled for a pre-generated video
  • Smoothly integrating interdependent components created by different teammates (so many merge conflicts…)

What’s next for Ming Management Training:

  • Talk and test with stakeholders
  • Scaling to learn patterns among employees in a specific company culture
  • Multi-conversational-agent scenarios, e.g. resolving a conflict between two coworkers
  • Ultra-low latency agent interactions
  • Immersive AR/VR interactions
  • Seamless mobile support

Built With

Share this project:

Updates