Blog
Articles, learnings, and thoughts.
Chips aren't the only bottleneck on abundant intelligence. 200 acres with a power path is the other.
Siting an AI data center is a search across 59 parameters. I built an agent harness to run that search, and made it report how much it doesn't know.
The age of the 10x engineer is ending. The 100x pipeline is next.
Coding agents are turning software development into a production-line problem. The durable advantage will be the pipeline, not the individual coder.
5G unblocked BVLOS in a weekend. The paperwork took eighteen months.
What I learned trialing 5G-enabled UTM and BVLOS drone operations: the network was never the bottleneck.
MARL trained the policy in weeks. Sim-to-real took the rest of the year.
Building multi-agent RL for tactical drone deconfliction in 2021. The training was the easy half — the simulator was the lie.
Federated airspace isn't a technical decision. It's a trust decision.
In 2019 we pivoted Vizzbee to a federated UTM architecture. The reason wasn't latency or scale. It was that no single party could be trusted to run the center.
The drone topped out at 0.5 m/s. None of it was the EKF's fault.
In 2018 we built drone autonomy on a T265, D435, Jetson Nano, and an EKF fusing VIO with GPS. The velocity ceiling wasn't where I expected it.