Natural Language Processing Engineer
2020 - PRESENTNovetta Corp- Designed an NLP dialog system in Tensorflow 2.0 for non-technical users to create complex, structured queries. Achieved near-SOTA NLP performance on limited dataset (>95% F1-score on 6 NLU tasks).
- Fine-tuned BERT with multi-task LSTM head models, multiple NLP input streams, self-attention. Substantial performance gains via augmentation, custom loss functions, CRFs, keras-tuner.
- Coordinated dataset design, collection, cleaning and annotation with six annotators.
- Developed an open-source solution for rapid, hotkey-based, multi-label annotations: https://github.com/etotheipi/keynotate.
- Mentored team members on machine learning, TensorFlow, AI architectures, Git, and Docker.
Technologies: Custom BERT, LSTM, Docker, Scikit-learn, Python 3, Keras, TensorFlowData Scientist | Software Engineer
2016 - 2019IronNet Cybersecurity- Engineered crucial features to improve a DNS-tunneling detection algorithm which reduced false positives by 80% and increased computational efficiency by 70%.
- Collected and analyzed data for detecting malicious meek (domain fronting) connections originating in customer networks.
- Developed a domain-generation attack (DGA) detection algorithm using a variety of machine learning (ML) techniques, including LSTM for identifying randomly generated domains.
- Served as the technical lead in the migration of the entire company product’s back end to Docker and Kubernetes (50+ microservices).
- Developed scripts for full-stack rollouts in AWS, including TLS certifications, DNS routes, security groups, and so on; also integrated them into the CI/CD pipeline.
Technologies: Amazon Web Services (AWS), Kubernetes, Docker, Apache Kafka, Machine Learning, Python, Spark, ScalaCEO | Lead Developer
2013 - 2016Armory Technologies, Inc.- Developed an open-source Bitcoin wallet from scratch that innovated multiple usability and security features in the early days of Bitcoin.
- It was a spare-time project for two years before receiving seed funding at a $4.2 million valuation in 2013.
- Became a respected thought leader in the industry, and the software still protects over $2 billion (USD) worth of Bitcoin.
- Managed five, full-time, remote developers.
Technologies: Cryptography, NoSQL, PyQt, User Experience (UX), User Interface (UI), Bitcoin, C++, PythonPhysicist, Computer Vision for Missile Defense
2006 - 2013JHU Applied Physics Laboratory- Worked on algorithms for "Lethal Aimpoint," which uses an IR camera on the interceptor missile to detect the threat and identify where to hit it. Used a variety of image processing and statistical techniques to solve the problem (missile defense).
- Developed an elaborate visualization tool that was used daily by dozens of engineers to examine and verify the results of missile simulations.
- Created a CUDA/C++ algorithm to speed up image-processing tasks in our simulations by a factor of 50-200x. The simulations originally took three hours to run, reduced to a couple of minutes with this code (and installing GPUs in our clusters).
- Developed statistical techniques to mitigate the effects of dead pixels in our IR cameras, and drive the requirements process for future camera/array production.
Technologies: Video Processing, Image Processing, MATLAB, C++