Led design and implementation of decentralized finance (DeFi) app facilitating automatic payments on Ethereum smart contracts. Built back-end API servers in Node.js with Metamask signature authentication as convenient interfaces with smart contracts. Designed front-end websites in React that pull and aggregate data from smart contract event logs with automatic synchronization. Integrated DeFi app with SAP’s Self-Sovereign Identity (SSI) project, an existing solution for secure digital identities. Led SSI development and on-boarded new software engineers as all original software engineers on the SSI team off-boarded.

Built robust back-end service in Golang to receive up to ~2000 requests/s about Captcha challenges across tens of millions of users. Designed tech spec from scratch. Enhanced security by actively designing against malicious users and enforcing HMAC authentication. Reduced maintenance efforts by 80% through using Kibana, Prometheus and Grafana to log and visualize service health metrics. Engineered ETL pipeline using Airflow DAGs to process millions of data points ingested by the service for data analysis.

Improved object detection performance of government-deployed YOLOv3 model by 20%. Built an ETL data preparation pipeline to automate data extraction and annotation, effectively cutting time needed for data preparation by 90%. Expanded feasible hyperparameter search space of DeepSORT by >20 times through Bayesian optimization. Built Nvidia-based Docker images to operate pipelines on AWS and Kubeflow. Overhauled monolithic legacy code into an end-to-end modular multiprocessing pipeline.

Publication: Skill Discovery for Exploration and Planning using Deep Skill Graphs, ICML 2020 LifelongML

Engaged in reinforcement learning research using Deep Skill Graphs to reduce large continuous MDPs into small discrete MDPs. Researched methods like Local Graph Partitioning to aid agent exploration in complex environments, eg. Montezuma’s Revenge. Currently extending Relative Novelty to large, continuous state spaces using Random Network Distillation.

Engineered DataQuality framework to efficiently identify malformed data in a relational database with millions of health records. Designed and built front-end for DataQuality framework, optimizing for client use and performance. Reduced development time of DataQuality framework by 85% through automation, greatly reducing development costs.

Publication: A Brain-Computer Interface (BCI) to Detect Responses to Affective Audiovisual Stimuli from EEG, IRC-SET 2017

Developed a BCI to control a LEGO robot wheelchair using electroencephalogram (EEG) and motor imagery. Architected a BCI to detect responses to affective audiovisual stimuli from EEG to aid patients suffering from alexithymia. Implemented front-end C# GUI program to run experiments and collect EEG for analysis from 29 participants. Analyzed experiment data using MATLAB to perform k-fold cross validation and calculate statistical significance of results.