Staff software engineer As the head of perception engineering team at Kodiak, I oversee the development of advanced sensor fusion and tracking systems for the Kodiak Driver. We take raw sensor data and create a machine and human interpretable representation of the environment that enables safe and reliable autonomous operations. We leverage both learned and classical approaches to accomplish this goal. Join our awesome team and help launch autonomous trucks!

Graduate Research Assistant - Stanford Intelligent Systems Laboratory, working on risk characterization of human drivers, broadly categorized as Imitation and Prediction.* Generative Adversarial Imitation Learning to create state-of-the-art human driver models. - Parameter Sharing to extend to multi-agent application (IROS 2018 - https://arxiv.org/abs/1803.01044, https://ieeexplore.ieee.org/document/8593758). - Reward Augmentation to significantly improve performance on emergent properties (ICRA 2019, https://arxiv.org/pdf/1903.05766v1.pdf, https://ieeexplore.ieee.org/document/8793750). - https://github.com/sisl/ngsim_env* Exploration into the use of Coupled Hidden Markov Models to leverage correlated data (no paper).* Developed real-time risk prediction system (from monocular video) that uses a webcam and can be deployed in any car. General: observe the current state of the world and simulate forward in time to develop an estimate of the risk to the driver over the next many seconds (5-20). - Includes a variety of technology, including: - FasterRCNN (deep convolutional net) trained on the Kitti dataset, - particle filters for object tracking, - advanced state estimation techniques, - advanced rules-based driver models for simulation, and - simple feed forward networks for risk prediction.

Worked on the perception team: developed, improved, and adapted advanced deep learning object detection networks to run in real-time, with a focus on performance in the safety critical automotive environment (TensorFlow, Python, C++).Ownership over design and implementation of tools automating data collection and curation, and model testing (Python, C++).Thrived in the fast-paced and multidisciplinary environment of a hyper-growth division. https://www.lyft.com/self-driving-vehicles

Implemented a feature for candidate recommendation system to expose model reasoning to system users.Built a system to analyze our logistic regression model and effectively extract important features (C++, Python, SQL).Displayed the most significant model features to users in human-readable form (Python, SQL, GraphQL, React, PHP).Analyzed the model to improve model calibration by over 10% and further improve prediction confidence (Python).

Implemented Long Short Term Memory Recurrent neural networks (LSTM) using TensorFlow to improve driver intention prediction.Compared classification performance of LSTM, Multilayer Perceptron, and a Bayesian Network (used Python and Julia).Published to IEEE Intelligent Vehicles Symposium 2017 (June), with 20-minute lecture presentation. - http://ieeexplore.ieee.org/document/7995948/ - for code in desperate need of a cleanup: https://github.com/djp42/IntentionPrediction

Detected anomalous foreign language xml files utilizing limited natural language processing in Python. Assisted in the set-up of 3 resource-heavy VMs (C) on a company-wide "cloud".Involved significant diagnostics and problem isolation and solving, and comprehensive documentation.

Responsible for designing communications link for a CubeSat that will travel to Jupiter’s moon, Europa. Created communications link budget, 1 of 2 undergrads on team of ~10. Stanford Space Environment and Satellite Systems Lab.Paper where I was a contributing author: http://www.sciencedirect.com/science/article/pii/S0094576516308797