Allen Gary Grimm

Built a high-quality spam filter for websites based on logistic regression and several iterations of feature engineering centered entirely on linguistic features. Developed image and video categorization using AWS Rekognition to extract keywords from media. Rewrote a DynamoDB-backed CloudSearch implementation wrapped in Lambda. Started from a failed proof of concept. Compiled and configured constraint satisfaction code (SCIP) for a client's use case. Wrapped software calls in a Docker container and deployed it as a service using AWS Batch. Built a custom recommender system—implicit user-item collaborative filter customized to return relevant people based on a product. Enclosed it in a Django project using DRF to serve as API powering a contractor-made web interface. Created OCR on Egyptian Hieratic using convolutional neural networks. Performed several web scraping with a simple GNU Wget and a bot that dynamically navigated the site. Debugged, updated, and cleaned inherited Looker Integration, now a part of Google Cloud, to enable both internal and external analytics dashboards. Implemented keyword extraction from resumes and job postings using TF-IDF and word2vec.

Optimized the PySpark-based uplift model from a runtime of eight hours down to five minutes on a benchmark dataset with millions of rows and hundreds of variables. Packaged the Uplift model into a properly versioned pip-installable package shared across the team. Added MLflow interface to uplift model to fit within the rest of the team's model pipeline. Updated the tree-based uplift model's decision functions to the cutting edge, increasing average model performance by 50 BPS across the team. Helped refactor the experimentation pipeline to better use Airflow and PySpark by bringing poorly scaling pipelines back within SLA requirements.

Led the processes of scoping and selecting possible machine learning uses, prototyping chosen initiatives, and productizing final models. Contributed to the development of the project Canonicalization. The core of Doing’s data is scraped event postings from several major event publishers. Through this, we frequently encountered duplicate locations across sources and duplicate events across and within sources. A distance-based test done theoretically comparing every event to every other event (but optimized enough to be computationally feasible; almost fast) or every location to every other location let us find events and locations that were so similar they were likely the same. This project was built from scratch up through productization. Helped build a tag extraction project. To help users quickly understand events, it is useful to have a short list of potent tags attached to each event. This project was prototyped using an aggregation of Doc2vec and Tf-idf. It was validated through systematically generating surveys via Google Docs to let the team give feedback on the quality of tags generated. Helped build a categorization project. Similar to tags, categories are useful to help us better understand our events and to help users better navigate available events. This was also prototyped using Doc2vec comparing each event to a whitelist of available categories (which came from picking the most popular categories listed by our data sources). This one reached the stage of prototype. Contributed to the development of the project DoingRank. Given a complete lack of user data (the startup’s app is still unreleased) but significant event data, none of the supervised recommender algorithms fit. So the first version (that only barely reached the stage of prototype) had two components. The first, to encode an abstract notion of event quality, was a math-ized version of the collective intuition of properties expected in good event postings (a title that matches the description, consistent event postings, etc). The second part is user-specific and maps RSVPs and other direct-app interactions through tags/categories to form a high-level notion of preference.

Conceived, prototyped, and productized data science initiatives. Researched models and wrote the valuable ones into the app. Created a relevance score model applied to content based on how users consume and react to content. It was a mathematical equivalent to a neural network, though the training was mainly done by interviewing domain experts due to little available data. Developed a model that generates tags attached to content based on who consumes what content in which context. For example, if many salespeople consume a document and nobody else touches it, the content is probably for salespeople. Documented and identified holes in current client-facing reporting infrastructure. Built new reports into the app as appropriate. My contribution mainly focused on the back end, but occasionally required front-end work too. Upgraded the current search engine to include spellcheck, faceting on our current tag infrastructure, and autocomplete.

Surveyed time series prediction methods. Conducted a case study on time series prediction applied to server usage in R. Wrote product-quality implementation of the chosen time series model (holt winters) from scratch in Python. Calibrated forecasting intervals (expected accuracy on predictions) in terms of performance, and trained and tested sets of data. Documented model implementation and testing procedures to enable the client's engineering team to build the model into their dashboard.

Modeled user activity and interactions to optimize the user experience by filtering content to what is likely to be the most interesting and useful. Helped build out back-end data infrastructure to improve app performance and prepare for scalability. Conducted A/B studies to help with product decisions. Clustered user behavior into distinct and comprehensible segments. Conducted and internally published the app's virality to report product success and direct product decisions.

Modeled and predicted user behavior in mobile games. Core projects included churn prediction and user path prediction. Managed relations between data science and engineering to catalyze productization of initiatives. Conducted ad hoc advanced analytics to assist in product decisions and to seed ideas for future data modeling. Rebuilt system logs: Solved for errors in observed device identifiers and marked invalid log entries as such. More precisely, the task was to write an iterative mapreduce algorithm to solve for all connected components in a several-billion node network using Hadoop Streaming and Python. Recruited, trained, and managed small teams of interns to assist with projects

Demoed data mining. Defined roles for new full-time data miners in a lab. Created a database architecture to centralize the lab's data collection and analysis. Worked with researchers to import their personal research data into a consistent format. Liaised with lab researchers and the Wolfram team to build the centralized database.

Built an evolutionary algorithm in C++ using the library ParadisEO to evolve complex networks. Wrote a network evaluation utility to simulate traffic and calculate other metrics on networks representing massively parallel processors with non-traditional interconnections. Built out and documented the experimentation process to enable fellow researchers within and outside of the university to use my framework. Conducted experiments relating the properties of links to the types of networks it would optimally be used in. Wrote a thesis on creation of a framework and the results of initial experiments