CodeAI (MeetUp)

This lecture presents a hybrid approach of software synthesis that combines Large Language Models (LLMs) with evolutionary algorithms to create software for specific domain. It covers the core theoretical ideas behind the approach and demonstrates how these techniques are applied in practice to build and develop software in everyday development workflows.

A year ago, we built an agent. Then we made it agentic. Now it's a full-blown multi-agent system. Remember when products didn't change under your feet every five minutes? I do. It was nice.

In this talk, I'll share Qodo's journey through AI's chaos — and what it means to benchmark a system that keeps reinventing itself. As our architecture evolved from single prompts to agentic flows to a full multi-agent code review system built on LangGraph and MCPs, our benchmarks had to evolve with it. I'll walk through how we approached evaluation at each stage, what broke along the way, and the hard-won lessons of staying rigorous in a market that won't slow down.

At high enough volume, you stop thinking about prompts and start thinking about systems. Over the past months, I've used Claude Code to manage complex, multi-stage engineering pipelines, batch processing hundreds of artifacts through generation, validation, hardening, and review cycles. This talk covers what I've learned about using Claude Code as a production orchestration engine.