top of page
normal crop.jpg


I am a principal researcher at Microsoft Research in Redmond, WA. I work in the Artificial Intelligence section of the lab, concentrating mostly on understanding and predicting peoples' location. This has applications for mobile computing, local search, maps, and driving. I earned my PhD in Robotics from the School of Computer Science at Carnegie Mellon University.


I gave a keynote talk at Ford's internal AI conference in June 2018. I explained our recent work on traffic modeling with a Markov random field, computing the value of GPS data, and finding safer driving routes.

I got a 10-Year Impact Award at the 2017 UbiComp Conference for a research paper I wrote in 2007 on location privacy called "Inference Attacks on Location Tracks".

My  coauthors and I got the best paper award at the 6th International Workshop on Pervasive Urban Applications, collocated with UbiComp 2017. The paper describes an algorithm for predicting the number of taxi pickups in New York City.


What are the best questions a personal digital assistant should ask to learn the most about you? We used mutual information on census data to compute the best few questions. This was work with Nikola Banovic when he was an intern with me at Microsoft Research.

Events induce people to tweet. Here is the average number of tweets sent from some NFL football games in 2017. There is a steep rise in number before the game starts, and then a gradual decrease. There are parallels to evoked mental responses in the brain.

safe driving route, Risk-Aware Planning: Methods and Case Study on Safe Driving Routes, John Krumm

Instead of computing the fastest driving route (green), we can compute the safest route (red), or a compromise between the two (black).

Eyewitness: Identifying Local Events via Space-Time Signals in Twitter Feeds, John Krumm

We can detect sudden spikes in the number of geotagged tweets from a region and find eyewitness accounts of what is happening before the news media has a chance to report it.

Mad Libs is a fill-in-the-blank game, where the goal is to create a humorous paragraph. We trained a machine learning model to help select funny words to fill in the blanks, and our judges said our computer-aided Mad Libs were funnier than the ones filled in by people with no help.

bottom of page