Mark McCurry

Developing semantic segmentation models for interior spaces Building infrastructure for managing ML data at scale

Researching methods for estimating trajectories of vehicles and pedestrians for use in an autonomous vehicle stack.

Developed algorithms for fingerprint matching, facial antispoofing, and voice antispoofing tasks • Hands on involvement from data acquisition, manual & semi-automated labeling, model development (Python/tensorflow), and SDK integration (C++). • Reduced error rates of facial antispoofing task from 10% to 3% with customized DNN and oversaw development of unique large scale internal dataset. • Reduced error rates of voice antispoofing task from 8% to 3% • Assisted in achieving MINEX III high performance fingerprint engine certification in C/C++ codebase • Delivered documentation for reproducible tensorflow models preprocessing, training, and deployment with jupyter notebooks • Worked with sales department to provide rapid updates to ML challenges * Developed PyTorch based models for classification, metric learning, and image enhancement tasks * Built docker based containerized workflow for training and testing models. * Lead shift to Triton for model serving

For this position I have been the developer of the Zyn-Fusion project. This project has involved the creation of a new 25 kloc user interface for the open source ZynAddSubFX (http://zynaddsubfx.sf.net ) musical synthesizer engine. This project has involved cross platform development of a new user interface toolkit using a combination of C, C++, and ruby. The sales of this product has helped development of the open source core of ZynAddSubFX (75 kloc of C++) which I have been the lead project maintainer since 2009.

At Georgia Tech I worked on state of the art signal processing and machine learning research. My research originally targeted speech processing and it broadened into the area of biosignal processing with a focus on computational neuroscience. While at Georgia Tech I worked on findings new machine learning and denoising options to make processing of EEG and LFP (local field potential) recordings easier and more reliable.

At Polycom I worked on embedded audio processing technology for use in conference calls systems. I focused on applying signal processing and programing knowledge to improve existing systems. This included optimizing an echo cancellation setup, prototyping acoustic room characterization recording hardware, and developing a configuration tool for embedded-software DSP graphs.

At this position I developed the hardware and software for a microphone array. This microphone array was then used in experiments to quantify the difficulty in the speaker recognition task when the speaker is at a distance and a beamforming array is used to record their speech.

At the MIT Haystack I worked on manipulating radio astronomy data using CUDA. The goal of the project was to produce a GPU/CUDA based solution to compress VLBI (Very Long Baseline Interferometry) streaming recordings before they would be written to hard disk packs. The delivered solution yielded roughly 70% of the performance of the pre-existing FPGA based system which indicated within a few years it would be practical to move to lower cost GPU based setups.