Autonomous flight.— from algorithms to hardware that ships.
I'm a robotics engineer obsessed with the moment a planner becomes a trajectory and a trajectory becomes a real machine moving through the real world. I work across the full stack — planning, control, embedded— because robots don't care which layer your bug is in.
Recently finished my MS in Mechanical Engineering at Carnegie Mellon, with a concentration in Robotics & Controls, where I worked on Hierarchical reinforcement learning at AirLab and hardware bring-up at CeRLab. Currently with the Learning and Controls Group on state estimation and Vision-Language-Action (VLA) models on Bayesian neural networks.
Systems that flew, drove, or grasped.
From-scratch C ADMM solver exploiting MPC structure via cached Riccati recursion — 63 µs median solve (13× faster than OSQP), 3.7 mm figure-8 tracking. Benchmarked against OSQP, a hand-rolled interior-point solver, and ReLU-QP, then distilled via DAgger into a 2.1 µs zero-dependency C++ neural controller.
Single policy unifies stand-up, wheeled traversal, and bicopter flight with no mode-switching logic — 3× tighter tracking than a tuned PID baseline. A hierarchical raycaster planner over a frozen tracker reaches 90% of goals with 4% collisions in clutter, deployed in ROS 2.
End-to-end Graph-of-Convex-Sets plugin for MoveIt2 with IRIS-ZO region growth, SOCP relaxation via CVXPY, and C¹ Bézier trajectory optimization. Caching brings repeat queries from 28 s to 2.4 s. Across 20 cluttered scenes, farthest-point IRIS seeding cuts path cost 9.8% over random — beating a PointNet regressor and a conditional VAE.
Currently working with the Learning and Controls Group at CMU on Vision-Language-Action models on Bayesian neural networks — combining language-conditioned policies with uncertainty quantification for safer robot behavior. Also continuing the DoubleBee bicopter-wheels work from AirLab independently, and on the CeRLab side, helping bring up the sensor stack and hardware for an autonomous wheel loader.
Open to full-time roles in motion planning, controls, robotics software, and mechatronics.
Last updated April 2026
MS in Mechanical Engineering, concentration in Robotics & Controls · GPA 3.90 / 4.00
State estimation and Vision-Language-Action models on Bayesian neural networks.
Hierarchical reinforcement learning on a hybrid bicopter-wheels robot.
Sensor stack and hardware bring-up for an autonomous wheel loader.
Consultant B1 / Assistant Manager
Applied ML and analytics to fraud detection and risk pattern identification across financial portfolios.
Junior CAD Engineer (Internship)
Gravity-load analysis of gas turbine engine components — HyperMesh + Nastran + LS-Dyna.
B.Tech in Mechanical Engineering
Bachelor's thesis: cloud-top-temperature-driven cyclone track forecasting via machine learning.
Click anywhere on the canvas. The drone uses an artificial potential field to skim around obstacles, and an arrive behavior so it settles cleanly on the target instead of overshooting. Toggle show forces to see what the controller is feeling.