Implemented the system API gateway using FastAPI and deployed using Nomad, passing stress tests with over 20k requests

Designed to handle over 40k queries per minute using ElasticSearch, MySQL, Redis and RabbitMQ, and built a web front-end for visualization

Built a license usage prediction system with an ensemble model of XGBoost regressors and LSTMs with a periodic accuracy monitor, mean absolute error as low as 0.39%; querying/posting data from/to Graphite and ElasticSearch, served using Nomad

Authored a 44-page write-up on matrix calculus with exercises and full solutions, selected as one of the recommended extra readings for the course

Participated actively in Piazza Q&A board with 1,309 total contributions; highest among all CMU MLD TAs in history

Hosted office hours on topics including neural network, reinforcement learning, logistic regression, recommender system and Bayesian network

Wrote/graded homework and exam questions on reinforcement learning and logistic regression

Provided starter code and set up Gradescope autograder for programming assignments

Achieved up to a 32.9% recall score improvement in interactive image retrieval by devising a neural network with cycle consistency

Designed a novel neural network for viewpoint-agnostic change captioning and attained an improvement of up to 15 points in common natural language processing metrics such as CIDEr

Published papers at AAAI 2021 and ICCV 2021, and gave an hour long lab presentation on recent image retrieval works

Collaborated with Hyundai Motor Company and completed funded research projects (value of approximately 240k USD)

Implemented neural networks with PyTorch, and created a dataset for change captioning under viewpoint changes using Blender