Presented some real-world learnings from building AskPrisma.ai using Autogen at the Applied AI Bootcamp: Architecting Production Ready Systems.

I had a great time presenting my work and learnings. It was a hands-on, operator-focused event where people building in production shared code snippets, edge cases, and lessons learned from interfacing with real LLMs in live systems.

What I shared

I presented a talk on “Lessons from Building a Multi-Agent System”, drawing directly from my learnings on AskPrisma.ai and how I’ve used AutoGen to structure a modular, multi-agent system.

The presentation covers:

  • What an “agent” actually is in practice
  • Considerations while building agentic systems
  • Design constraints I ran into: context windows, orchestration, persistence, debugging
  • How AutoGen’s abstractions and agent presets helped
  • Discussion on tradeoffs

Link to the presentation: Lessons from Building a Multi-Agent System

Other highlights from the bootcamp

  • @getpy spoke on DSPyOSS usage for processing real estate call centre transcripts and a DSPy crash course
  • @_anks and @abhicantdraw shared practical insights around evals and keeping systems simple — including examples from their work at ClearTax
  • Girish from Wefab showed a jaw-dropping demo of how just two engineers rebuilt a full vertical in 2 months using the frappetech framework, config-driven UI, and an MCP server

Final thoughts

It was great connecting with others thinking deeply about agentic systems, multi-step reasoning, and what real LLM product architecture looks like when the rubber meets the road. Props to @getpy for bringing together a thoughtful group and a solid lineup of talks. Special thanks to @_anks, @abhicantdraw, and Girish for sharing lessons that were equal parts tactical and motivating.

Looking forward to more such spaces where builders get to share what’s real.