Built a Python stitcher (PyTorch, OpenCV, XFeat, RANSAC) that merges out-of-order, GPS-tagged photos from 3 drones into one georeferenced map. Sole developer, merged with 130 tests.
Designed a least-squares pose refinement that cut mean simulated misalignment from 13.6–69.8 cm to 4.0–8.3 cm.
Solved feature matching across 180° row turns, then used drone yaw to extract features once per photo, cutting simulated time per photo from 0.42–0.75 s to 0.13–0.19 s.
Implemented lens undistortion and blur, exposure and tilt checks that keep unusable photos out of the map.
Built a swarm-spacing module (Lloyd’s algorithm, Voronoi cells, Shapely) that spreads drones evenly over concave, GPS-bounded fields, cutting spacing variation (CV) from 0.46 to 0.06 in simulation.
Homework Helper and Assistant Teacher
September 2022 to July 2025
Spirit of Math Schools Inc., Richmond Hill, ON
Ran a drop-in homework classroom as one of two staff (2024–2025), helping about 30 students per session.
Taught 1–2 classes a week of up to 16 students (2022–2024), covering lessons and contest-style algebra.
Developed a Pico 2 andon system that hears a fault through an I2S mic and cuts the fan’s power via a MOSFET, stopping each of 3 real fans by itself 3.461–11.661 s after an injected fault.
Wrote real-time C firmware (I2S capture via PIO and DMA, FFT log-mel features, per-fan Gaussian mixture scoring) that matches Python’s scores on the board at 1.684 ms per 32 ms audio frame.
Trained a PyTorch CNN on healthy audio only, reaching 92.13% mean AUC on held-out DCASE 2020 data against 79.70% for the baseline autoencoder I reproduced.
Proved the model hears more than loudness: with volume matched it still caught both real-fan faults at 100% AUC, while a level meter reached only 63.34% AUC on the DCASE dev set.
Quantized the CNN to int8 (383.0 KB to 98.7 KB at 92.08% AUC) and ported it to C, bit-exact with Python on the Pico 2’s Cortex-M33.