On-device knee moment inference
A ~90k-parameter knee moment model I trained, running in about 2 ms on the Teensy 4.1 that runs the exoskeleton's control loop. On the bench so far.
Experience
Internships and research. The boards I designed at work are confidential, so the photos here are my own bench shots or public ones, and the pages say what I did in words and numbers.
Work and research
CMU MetaMobility Lab
Graduate Researcher · Pittsburgh · since September 2026
The MetaMobility Lab at Carnegie Mellon builds lower-limb exoskeletons for the hip, knee and ankle. I work on sensing hardware, firmware and on-device ML inference.
metamobility.cmu.eduTornyol (YC F25)
Hardware Engineering Intern · Paris · May to August 2026
Tornyol is a seed-stage startup building small autonomous drones that intercept mosquitoes. I worked on the drone's electronics for the summer. The boards are confidential, so the photos here are Tornyol's public ones, not my boards.
tornyol.com
22 × 22 mm drone flight controller
A 6-layer Betaflight flight controller around an STM32G473, with four custom ESCs, that runs on a LiPo or on supercapacitors.
20 A supercapacitor charger
A BQ25820 charger with an op-amp cell balancer and my STM32 firmware. Ran 20 A on a 1S pack and charged the 2S flight pack at 12 A.
VL53L9CX ToF bring-up and lidar flight model
Brought up a VL53L9CX ToF sensor on an STM32N6 and retrained a GRU flight model's input layers for it: 90% obstacle avoidance in simulation.
Withings
Two internships · Paris · September 2024 to July 2025
Withings makes connected health devices: watches, scales, sleep trackers and an at-home urine lab. I interned there twice, first in industrialization on a production line, then in R&D on a smartwatch.
withings.com
Low-power snoring detection on a smartwatch
Audio ADC drivers and a duty-cycled wake-on-sound pipeline that cut the feature's battery cost from 15 days to 1.7 on a 30-day watch.
Production-line vision for U-Scan
Python and OpenCV defect checks on a Raspberry Pi station, deployed on the production line in China, then a TensorFlow Lite autoencoder back in Paris.