Director of Engineering at PodFirst, where I lead delivery on PodCut and GoBox. Before this, two years building intelligence into humanoid robots. The thread through both is the same: getting models to hold up in systems that have to actually work.
PodFirst's proprietary AI post-production engine, with a web app and a Premiere Pro integration. I run execution between the founders and the build team, and own the work of bringing several AI models into one editing pipeline. Editing that used to take a studio team hours comes down to minutes of supervised output.
A portable studio you can unfold in a hotel room or an empty office and record broadcast-quality audio and video. We're moving from working prototype to first production run. Mostly that's a fight with suppliers, tolerances, and unit cost.
A podcast on legged robots, hobby builds, and the side of hardware development that doesn't make it into papers. The conversations that happen in Discord channels at midnight rather than in formal venues.
I build intelligence into machines. For two years that meant humanoid robots: training neural policies for legged control, working out why an actuator that behaves in simulation falls apart on real hardware. Now it means software at PodFirst, where I lead delivery on PodCut and GoBox.
The work looks different from the outside. From the inside it's the same problem. The model is rarely what makes or breaks a system. Everything around the model is. A controller that holds up in simulation and fails the moment it meets a real actuator is the same failure as an AI feature that demos well and falls over on the fifth client job of the day.
Most production AI I see fails the way humanoid robots fail in simulation. The discipline that fixes it isn't a better model. It's the habit hardware engineers have of identifying the system before they trust it. That's the thread I work on, in robotics and in software, at the same time.
Before PodFirst, a few years in robotics R&D and manufacturing automation.
Designed in Fusion 360 and physically integrated: a Universal Robots UR-10 collaborative arm on an Omron LD250 autonomous mobile base, inside a custom frame that enclosed the power systems and the robot controller, with a SmartShift tool changer for swapping end effectors.
Took the same use cases onto humanoid platforms, the Unitree H1-2 and G1: training neural policies for legged control and working the sim-to-real gap on real hardware, where a controller that behaves in simulation comes apart against a real actuator.
Project manager on an eight-month build of a next-generation trainer, a compliant robotic mechanism that gives force feedback for training surgeons.
Working roboticists, hobbyists deep in a build, founders in legged or applied robotics. If that's you and you want to record, get in touch.
Pitch yourself as a guest →Building in hardware, putting AI into production, or up for a collaboration? Same inbox. {{ email }}