I’ve worked on a number of projects, most of them involved extracting training data from massive production logs using hadoop (MapReduce, Hive, Spark), examining and visualizing data (Jupyter) and training models using the extracted data (xgboost, LR, BERT). Some of the projects are:- Search query completion - retraining existing model from scratch (lift of 4% in CTR) and developing context-aware model- Brand-phish detection - evaluating an existing model and training a new BERT-based model. Training process utilized LLM for labeling training data. Offline evaluation has shown a twofold increase in coverage without negative impact on precision.- Ad arbitrage (GINSU) - analysis of ad performance and generation of ad titles using templates- Automatic detection of “feel-good” news- Better news ranking model - added features integrating user feedback on previously displayed news- News articles enrichment with related search queries - built a complete pipeline for building relevant queries’ db and for matching queries to news articlesI’ve assisted several interns in navigating Yahoo internal data sources and using Yahoo infrastructure as well as building the training data.I’ve contributed to writing papers (accepted to major conferences) and patents based on our projects (see below)

I led a team of software engineers and linguists that developed from scratch state-of-the-art multi-language keyboard for mobile phones (the linguistic part of it). Among other languages it supported Arabic, Vietnamese, Thai, Korean and Simplified Chinese. The keyboard had prediction, completion and swipe input method for alphabet-based languages. For Chinese it had pinyin input that supported completion, prediction and input by initials, like nh for nihao.We developed in C++ so our code ran unmodified on (millions) of both iOS and Android devices.

Developed smart keyboards - see above.

I developed both rules for detection and correction of grammar mistakes in English as well as infrastructure that supported writing the rules.In order to recognize long distance dependencies and punctuation mistakes that required understanding sentence structure we developed a novel semi-shallow parser that was robust in the face of natural unedited text that could have missing words, including function words, wrong agreement, run-on sentences (missing period) and so on.

Development of (statistical) methods for grammar checking and correction. Introduced good engineering practices such as code reviews, source control, automatic nightly builds with test coverage reports and so on.Developed internal tools to assist analysis of test results, e.g. smart diffs.

Verification department: Piparazzi - CSP-based test generator for verification of processor's micro-architecture

Various small web-related project (Java, JavaScript)GUI for Hebrew tagger (Java)

Optimization of virtual calls (C++) in xlC compier