Startup CTO & AI Builder

The hardest part of AI agents isn't prompting. It's observability. After building real multi-agent workflows with Gemini, Codex, and Claude Code, here's what I learned.

Most AI agents fail because they improvise. The Octonauts and Shelldon Swarms work because each agent has a clear role, limited permissions, and a defined place in the chain of command.

I used to be a startup CTO, and I still evaluate developer tools through one lens: do they reduce friction and help teams move? Here’s my practical take on Codex, Claude Code, Gemini CLI, Gemini Code Assist, and Jules—and why I still personally prefer Codex.

A technology leader's reflection on the recent shift in AI architecture—moving from vanity metric scaling to true omni-modal problem solving. We've spent the last few years treating text as the universal language of AI. It wasn't. It was just a limitation we had to accept.

A technology leader's 2025 reflection on growth beyond vanity metrics — rate of change, time to solve, presence at home and a focused sprint into 2026. He does not measure growth by net worth or follower counts. After some hard pivots, leadership stretch, a couple of hospital scares, and the reminder that presence at home matters more than any metric, he is closing the year grateful and clearer. Many doors have closed, better ones opened. In 2026, he is sprinting — building where it counts, governing his time, and choosing the right doors on purpose.

A CTO's deep dive into the engineering behind DeepSeek-V3 and R1. Exploring MoE, FP8 training, MLA, and how open-source AI is democratizing the market.