My Research and Development at Amazon
Causality from non-experimental data (proprietary)
My research is mostly proprietary. In general terms, I estimate the causal effect of making improvements to the catalog data to the key perfomance metrics such as page views and purchases. This is extraordinarily complex because of the number of subsystems that each affect customer experience, and feed off the catalog data. It is not classical A/B testing for reasons of scalability, we have to make do with observational data. That has all sorts of statistical validity issues that we need to mitigate. We built and maintain a system using a large scale data pipeline feeding an econometric model that does the estimation.
Design and analyze experiments to answer business questions, define sampling protocols to deal with wild data exhibiting power-laws distributions, ...
My Research and Development at Boeing
Before Amazon, I was an Advanced Technologist (research scientist in disguise) at Boeing Research and Technology in Bellevue, WA. These are the things I've done there. Due to proprietary nature of the projects, the descriptions are deliberately vague - sorry.
Large-scale sensor data analytics (proprietary)
How does one build self-service predictive data analytics for engineers who are not experts in computing, but rather the system where the data originates?
The XDATA project
Understanding Purposeful Behavior
Using methods of inverse reinforcement learning, computers actively learn from humas to really understand observed behavior (defined as: ascertain and interpret the incentives and beliefs that explain the behavior as rational) of large numbers of agents, creating a ISR data exploitation capability to concentrate analaysts' attention on unusual and suspicious behavior instances, alert and generate explanations of the observations.
DARPA Bootstrapped Learning (Phases 2 and 3)
The BL program attempts to implement the "Bootstrap Learning Dream", which is to dispense with the need for programming. Instead, the agent (such as a UAV) is taught how to perform the required tasks by somebody who understands the problem, instead of understanding programming.
Agent Executive based on Partial-Order Planning (proprietary)
Wrote a simple executive for an autonomous agent that uses the "repairability" of partial order plans to react in an environment where actions are nearly deterministic (so you can ignore uncertainty in planning), but robust recovery from action or resource failure is essential.
Automatic Derivation of Decision Policies (proprietary)
How to do reinforcement learning when rewards are large but rare. Automating reward shaping.
Computer Vision & Machine Learning applied to Satellite Imagery (proprietary)