Hi! I’m a software engineer who likes building things that don’t exist yet and solving hard problems with what I build. Right now I’m rethinking how AI systems are built around humans and how to push through the current bottlenecks.
I work across the stack, from backend to product.
Leading a deep learning project of a team of 6 to develop a end-to-end GeoGuessr location prediction interface using both classification and regression
•
Building new models using few-shot representation learning and Meta-RL, and improving existing models performances such as Google PlaNet CNNs
Worked with Launchpad internal team on Alexa voice conversion (VC) model that takes in any sample voice and converts voice assistance to the speaker’s voice
•
Feature engineered vector representations of model inputs with low-level Python libraries such as Pyworld, Pysptk, and speech synthesis libraries like nnmnkwii
•
Trained CNN models on audio files converted as MFCCs with filterbanks, Reduced training time with time-warped parallel data and Mel Cepstrum predictors
Automated generation of summary reports from iShares time series data, deployed into production within a month and achieve high evaluation metrics
•
Implemented few-shot ML models for natural language generation using dual weights LSTM, Hidden Markov Models, and PLM
•
Improved model scalability with customized GPT, eicient data generation, and Tableau API integration for augmented analysis and data insight derivations
•
Delivered a customizable NLG model which removed the need for a third party solution costing over $300k
