BUILDING

Remus is the first motor from NEAL Aerospace, the student rocketry group I co-founded at Georgia Tech. It's a hybrid — nitrous oxide fed from a ground-side tank, through a run valve, into a 3D-printed ABS fuel grain — and it isn't a launch vehicle. It lives bolted to a test stand: the oxidizer stays on the ground beside it, and the whole point is to fire it, instrument it, and understand it. We chose the hybrid architecture deliberately, too — a hybrid can be shut off, it can't detonate the way a solid can, and it forces a small team to get genuinely good at one hard thing: moving oxidizer from a tank to a combustion chamber, on command, without leaks.

My corner of the system is everything the oxidizer touches. The injector came first: a MATLAB sizing model trading mass flow rate against pressure drop to hold a 350 psi chamber, run over and over as trade studies until the numbers stopped moving. An injector that's too soft lets chamber pressure oscillations talk back to the feed system; too stiff and you're throwing away tank pressure you paid for. When the sizing settled, I took the chamber flow into Ansys Fluent to see whether the combustion picture matched the assumptions the model was built on.

The less glamorous half of the job is the P&ID — the single drawing that names every valve, regulator, transducer, and fitting between the fill cart and the injector face. Writing it teaches you a mindset: you aren't designing for the day everything works, you're designing for the day something doesn't, and the drawing is what tells you which valve saves you.

Remus is now moving through component procurement toward cold-flow and static-fire testing on Ostia, the stand we built for exactly this purpose.

For a semester on Georgia Tech's Autonomous Drones VIP team, I worked on the other end of aerospace: vehicles that decide for themselves. The project trained drones to do full 3D guidance, navigation, and control in a multi-agent simulation — pathfinding through space, coordinating with teammates, and engaging targets — using PPO reinforcement learning.

The agent starts knowing nothing. No physics, no waypoints, no notion of what a target is — just observations, actions, and a reward signal. Everything it ends up doing well is something the reward function taught it, and everything it does strangely is something the reward function accidentally permitted. Most of my real work was there: tuning rewards and watching agents exploit gaps I didn't know I'd left, then closing them. It is debugging, except the bug is in your own statement of what you wanted.

That loop — specify, train, watch it misbehave, respecify — steadily improved both convergence and the maneuvers themselves. It also changed how I think about the hardware side of my work. A feed system and a reward function fail the same way: not because the design was wrong, but because the requirements were incomplete. Writing better requirements is the actual skill, and I've practiced it in two very different languages.

The VIP semester also opened a door: the following spring I joined the Aerospace Systems Design Laboratory for AFRL-sponsored multi-agent ML research — a separate project on a much larger canvas, where I built the Common Research Model framework and scenario pipeline the lab's studies run on. That work is wrapped now, but the through-line from both — simulation is only as good as its match to reality — is exactly what Voron exists to close.