AI / ML portfolio

Most people met automata theory and immediately started questioning their life choices. I spent my time making sure students didn’t stay in that phase for long. Between office hours, exam prep, and problem-set walkthroughs, I turned dense definitions and scary proofs into something that actually felt navigable.

On the backend, I designed and graded assignments with surgical feedback so students knew exactly why something worked or broke. Regrade requests dropped once the rubrics got sharper, and I kept Piazza, grading queues, and logistics moving so the course felt less like chaos and more like a well-behaved DFA.

Attendance used to mean calling names in a crowded classroom. I helped replace that with a facial-recognition system that checks in 200+ students in under a second per face, clocking around 87% accuracy across a 500+ face gallery. Professors stopped asking ‘Who’s here?’ and started asking ‘When can we roll this out everywhere?’

Under the hood, I squeezed the encoding and inference pipeline until it behaved: latency dropped, reliability went up, and the React Native front-end stayed smooth. The demo to university leadership turned into a campus pilot and an IEEE publication, which is not a bad output for one internship.

Wind looks random until your forecast model starts getting judged on mean absolute error. I fine-tuned an LSTM for wind-speed prediction that beat the baseline by about 10%, which is the difference between ‘yeah, sort of’ and ‘this is actually useful.’

Then I wired in Flower to run federated learning across four clients so each site could train locally without shipping raw data to a central server. Scripts, configs, and experiment scaffolding were all packaged cleanly so future datasets can be dropped in without archaeology through half-broken notebooks.