Staff Software Engineer, Google
Elsewhere: blog
Last updated: Oct 2020
I joined Google out of graduate school.
At Google I worked on ranking, recommendation, structured data extraction, and data modeling to help users find delightful places in Google Maps and Travel, and relish fine passages in Google Books.
Most of my development was in C++ or Python or some variant of SQL augmented with frameworks like Apache Beam and pandas. My favourite tools are interactive environments that let me dig into data easily — Jupyter notebooks and (perhaps surprisingly) Google Sheets. Interactive tools let you explore large data sets (logs, web signals, reference data sets, code outputs) and give you evidence for the important questions:
- what might be interesting enough to users to build?
- what signals might be useful to build into something great?
- what the heck is wrong with maciej's code?
On Google Travel I ran efforts across Google's Seattle, Mountain View, New York and Zürich offices to make Google Search and Maps a great Travel Assistant.
| Google as Travel Assistant | ||
 |
 |
 |
| From trip details | to top sights | to dishes to try and bespoke recommendations. |
Contextualization through Email Understanding I led a team of great engineers to build the software Gmail uses to automatically extract flight and hotel reservations from free form e-mail. The challenges here were factoring the problem, putting together a good debugging framework, and keeping the code fast and contained enough to process billions of emails a day that we could never observe. Over my tenure we doubled the recall of the system to 90-95% and reduced our error rate from 20% down to 1.5%.
The extracted reservations enabled assistance by adding events to the user's calendar and view of Google Maps so their flight number, status, and hotel are highlighted when they need it and where they need it.
All the world's Tourist Attractions, ranked I was a team lead for the group that combined the best of the web with Google's Knowledge Graph to select which places should show up as attractions and decide how they're ranked in Google Search and Local Search, as well as how they rank and appear on the base layer of Google Maps. This work affect tens of millions of Search results each day, and led to millions more place sheet opens per day from Maps and Search as users more easily found the places that would delight.
Personalized Travel Recommendations Though world capitals have literally tens of thousands of tourist attractions, users don't tend to look past the first dozen. For users who opt in to share record their location with Google Maps Timeline, we give them bespoke travel recommendations. Will Bradbury and I launched this work to Google's Travel Mobile experience. Our recommender is based on an embedding model we developed for attractions combined with training per-user classifiers based on what the user choses to do — or not-do — on their vacations.
Iconic Local Dishes I augmented Google's travel information with dishes that a traveller ought to try when they go to a new country, leveraging the wisdom of the web and pools of local experts found through Google Opinion Rewards.
Local Search I ran various projects to improve results for queries like beaches near me and things to do in bali. Improvements to these results came from a variety of efforts: ontology hacking, fighting spam listings, working with Geo operators to improve Maps and Knowledge Graph data, and rewriting chunks of the Maps and Search rendering stacks to better utilize data we already had.
I worked on the (open source) Tesseract OCR engine which Google used to digitize 20 million of the world's books as part of Google Books, making millions of public domain library treasures searchable and downloadable.
My contributions included cutting Tesseract's character error rate on English by 5% using word bigram patterns, developing Tesseract's Paragraph detector and adding Bi-Directional OCR support as part of the Tesseract 3.0 release.
Google Checkout, the precursor to Google Pay, helped online merchants process billions of dollars of GMV. I helped launch Checkout, building out replication infrastructure and merchant features, diagnosing design issues, and building the team's production monitoring.
Google's Traffic SRE is responsible for the Terabits of network traffic entering and leaving Google's production network and edge serving infrastructure (DNS, HTTP terminators, loadbalancing and Denial of Service detection and response).
While on the team, I did a lot of the typical Software Engineer and SRE work: simplifying and automating processes, responding to novel failures and attacks and figuring out ways to prevent them in the future, and wherever possible rendering whatever it was I was doing obsolete.
I designed and ran production qualification for DNS rollouts -- including capacity planning, building failure simulation software, load testing, and regression testing. I deployed anycast DNS through Google's production clusters which consequently made "production maintenance" no longer take out a rotating set of ten thousand machines every week.
I qualified and rolled out Dave Presotto's proprietary load-balancing DNS
server as Google's primary user-facing DNS ns[1-4].google.com. This
replaced our use of BIND, and while I was at it I converted these VIPs to
anycast. This rollout enabled Google to loadbalance naked domains for the
first time, and as a consequence reduced user latency by an average of
75ms for 3% of HTTP queries to Google.
Patent: U.S. Patent US8196198B1 Access Using Images (2008)
Code: Tesseract OCR engine, The DC3 Suffix Array Creation Algorithm (an implementation) helpful for aligning OCR output with the correct text.
-
PhD Dropout (Programming Languages), Carnegie Mellon University, 2005
-
B.S. in Mathematics, Georgia Tech, 2003 (4.0 gpa)
-
B.S. in Computer Science, Georgia Tech, 2003 (4.0 gpa)
Graphics coursework under Jessica Hodgins and Compilers with Olin Shivers.
-
Budapest Semesters in Mathematics; Budapest, Hungary 2002
- Fulbright Scholar, Hungary 2003-2004
- NSF VIGRE Grant Recipient for Summer Research in Mathematics, 2002
- Albert C Jacobs, Phi '21 Award, 2001
- Psi Upsilon, Gamma Tau Chapter President, 2000
- ThinkQuest Gem Award, 1997
-
IBM Extreme Blue Internship with Austin Research Lab, 2001
Developed Linux port to and application software for the prototype PowerPC 405LP
- Wrote PCI and embedded drivers for an LCD Controller
- Optimized video-related software stack (Linux Framebuffer / Nano-X / FLTK)
-
HP (nee Convex), 2000 bringing up IA64 machines,
-
IBM Linux Technology Center fixing glibc PowerPC bugs,
-
Georgia Tech Animation Lab porting simulation code from sproc() and POSIX threads and integrating Perl as a scripting library with C++.