I build software that lets machines sense the physical world.
My work spans state estimation, embedded data acquisition, and GPU acceleration. I co-lead software for USC's autonomous underwater vehicle, write flight avionics at the Rocket Propulsion Lab, and study how robots interact with sand in the RoboLand lab.
Why this work.
Robots only act as well as they can tell where they are. I work on the layer that turns raw sensor readings into a state estimate a controller can trust: the firmware that samples without stalling, the estimator that fuses what the sensors report, and the kernels that make it fast enough to run on the vehicle.
On the USC Autonomous Underwater Vehicle team I co-lead software, including a ROS 2 and GTSAM state estimator that fuses IMU, DVL, depth, and stereo point clouds. At the USC Rocket Propulsion Lab I write C++ avionics on STM32, from DMA acquisition pipelines to sensor drivers. In the Feifei Qian lab, RoboLand, I build the data pipeline behind a model of how sand resists a robot's foot.
I also build in the open: FlashICP for GPU point-cloud registration on Jetson, and a steady stream of fixes to robotics, embedded, and inference projects upstream. My developer tools live on the developer page.
Training.
B.S. Computer Engineering and Computer Science
Dean's List, Spring 2026. CURVE undergraduate research fellow in the Feifei Qian lab (RoboLand).
Kaggle silver medal, Kaggriculture: 413th of 10,148 teams.
A.S. General Science, Highest Honors
Associate degree with Highest Honors.
Where I build.
Undergraduate Research Assistant, CURVE Fellow
- Built a Python pipeline that turns robot force measurements in sand into ground-truth granular stress coefficients, giving the lab an independent check of its terrain model (r = 0.88–0.98).
- Found a clock mismatch that ran filters 1.9× too long and a start-up transient that flipped force signs; fixing both raised sign-correct configurations from 1 of 48 to 45 of 48.
- Traced a roughly 5° angle bias to rig deflection under load, using a symmetry test that needs no reference map.
Software Co-Lead
- Architected a ROS 2 / GTSAM 6-DoF state estimator for the Barracuda vehicle, fusing IMU, DVL, depth, and ZED point clouds through IMU preintegration, custom factors, and relative-pose constraints.
- Rebuilt the ICP registration front end with KD-tree correspondence search, estimator-predicted warm starts, and quality gates; traced localization divergence through rosbag replay.
- Built a py_trees mission planner whose gate and slalom behaviors recompute waypoints from live detections.
Embedded Software Developer
- Co-developed a DMA-capable ADS131 acquisition pipeline on STM32 with CRC validation, cache-aligned buffers, and asynchronous SPI reads, so sampling never blocks the CPU.
- Wrote SPI drivers for IIM-42352 and IIS3DWBTR vibration sensors and a polymorphic sensor-validation framework on Mbed OS.
- Led board bring-up over SWD and bus debugging across I²C, SPI, UART, USB, and Ethernet; moved telemetry from UART to USB 2.0 links.
Learning Assistant, TAC 348: Making Smart Devices
- Support students in USC's project-based smart-devices course as they build connected devices on the Particle Photon 2: electronics, embedded C++, sensors, motors, and wireless communication.
- Help debug wiring, circuits, and firmware during labs and office hours, from a sensor reading nothing to a device that won't connect to the cloud.
- Guide students through final projects, and pass recurring sticking points back to the instructor.
Embedded Math Tutor
- Completed 268 hours as an embedded calculus tutor, developing targeted strategies for students from diverse backgrounds; every student I tutored passed.
- Tutored one-on-one, answered questions on discussion boards and by email, and gave professors feedback on student comprehension and engagement to guide their teaching.
Estimation that holds up outside the lab.
State estimation
Factor-graph fusion of inertial, acoustic, and visual sensors, and the failure modes that only show up in the water.
Point-cloud registration
Fast, robust ICP on embedded GPUs, with warm starts and quality gates so a bad match never reaches the estimator.
Embedded data acquisition
Firmware that samples at rate without stalling the CPU, and tooling to validate packets and links before hardware exists.
Robot–terrain interaction
Measuring how granular media resist a moving foot, and turning noisy force logs into models a robot can plan with.