Luke Simon

AI research for ad ranking, including: LLM retrieval, LLM ranking, generative retrieval, generative ranking, generative ad auctions, causal ranking, and reinforcement learning.

AI research applied to retrieval and ranking for recommender systems and search using partial evaluation and staged inference of LLMs, embedding and KV caching of bi-encoder LLMs, contextualized late interaction LLMs, hierarchical cross-encoder LLMs, and reasoning foundation model LLMs for recsys. Large scale recsys and auctions using LLMs. Co-designing AI models and AI infrastructure. Lots of GPU hours.

Head of Twitter ads machine learning, including ad targeting, embedding based retrieval, ad ranking, ad auction, machine learning, and data science driven revenue growth optimization. ML based identity bridging.

Ad targeting, embedding based retrieval, ad ranking, ad auction, recommender systems, machine learning, and artificial intelligence for revenue growth.

Operated as engineering lead of Twitter’s machine learning and data science department for advertising-based revenue generation. Built, hired, and led team of 75 engineers. Liaised with sales team and external clients. Influenced product strategy and product management. Led product development, engineering roadmap and execution, systems architecture, and experiment review.

Served as engineering lead of Twitter’s ads targeting and modeling team for advertising-based revenue generation. Built, hired, and led team of 15 engineers. Liaised with sales team and external clients. Influenced product strategy and product management. Oversaw product development, engineering roadmap and execution, systems architecture, and experiment review. • Leveraged machine learning to deliver tools that track users’ identities and infer user demographics and interests, enabling advertisers to reach niche interest groups and personas that resonate with their business. • Built models for classifying and detecting topics in text, image, and video content, allowing advertisers to specify the type of content they want their ads to be displayed with. • Utilized multiple modeling algorithms and techniques, including deep learning, GBDT, random forests, explore/exploit, active learning, factorization machines, multi-modal embeddings, and approximate nearest neighbor search using cutting-edge spatial index data-structures and algorithms.

Led systems architecture, programming, and software development. Trained ML models. • Created Twitter’s ad keyword targeting product, including developing a deep learning model for keyword broadmatch. • Developed Twitter’s gender-based and age-based ad targeting products, including developing ML models for predicting users’ genders with 90% accuracy and ages with 70% accuracy. • Created Twitter’s geolocation-based ad targeting product.

Served as engineering lead of Microsoft’s Bing Search whole page ranking team. Hired, built, and led team of engineers. Oversaw engineering roadmap and execution, systems architecture, and experiment review. Click prediction using machine learning. Led modeling of relevance estimation system for federated search. Oversaw design and implementation of infrastructure and machine-learned models for named entity resolution and local search results. Developed explore and exploit system for improving click prediction models.

Served as engineering lead of Microsoft’s Bing Search whole page ranking team. Hired, built, and led team of ten engineers. Oversaw engineering roadmap and execution, systems architecture, and experiment review. Designed and implemented configuration-driven system for specification of relevance signals for query-time click prediction using machine learning. Led modeling of relevance estimation system for federated search. Managed design and implementation of infrastructure and machine learned models for named entity resolution and local search results.

Led systems architecture, programming, and software development. Trained ML models. Implemented stochastic gradient boosted decision tree training toolkit in C++. Developed ML models for whole page ranking of Bing search results. Developed explore and exploit system for improving click prediction models.

Implemented, in C++ and Java, a static program analysis for detecting all possible thread races in Java applications. Designed and prototyped, in C++, a new algorithm for improving the scalability of whole-program pointer analysis. Implemented, using C++, Java, and Datalog, a static analysis of data- and control-flow dependencies in Java applications, and lead a team of engineers in the development of a tool for navigating graphical diagrams of application dependency maps. Designed and developed a language-agnostic program slicer, in C and Java, for automated impact analysis, program dependency analysis, and version integration interference analysis.

Hired, built, and led team of seven engineers. Oversaw product development, engineering roadmap and execution, systems architecture, and experiment review. Designed and developed language-agnostic program slicer for automated impact analysis, program dependency analysis, and version integration interference analysis.

Implemented a static program analysis for detecting all possible thread races in Java applications. Designed and prototyped a new algorithm for improving scalability of whole-program pointer analysis. Implemented a static analysis of data- and control-flow dependencies in Java applications, and led team of engineers in development of a tool for navigating graphical diagrams of application dependency maps.

Researched reinforcement learning, static program analysis, verification, programming language semantics, logic programming, compilers, and operating systems. Published technical papers in peer reviewed scientific journals and conferences, and taught computer science classes. • Led research and development of new semantic markup language for web services called “Universal Service Description Language”, implemented using OWL, Prolog, and C. • Researched, designed, and implemented, using Prolog and C, a specification language for real-time problem domains for automating synthesis of plans for accomplishing goals by executing real world actions in a given situation. • Researched, designed, and implemented, using Prolog, a next generation general purpose logic programming language for reasoning about infinite structures and processes, with applications to model checking and systems verification. • Managed the development of a web service for tracking chemical inventories of factories, plants, and warehouses throughout USA, as well as hazmat emergency and medical treatment information for an emergency preparedness and response search engine.

Developed web based chat room with pure HTML front-end. Implemented a Java applet high volume chat room. Developed a Mac/PC interactive multimedia CDROM using the Lingo scripting language and Macromedia Director. Using Flash, developed an animated splash page for client’s web site.

Designed and developed graphical dataflow diagramming tool, using Visual Basic and C++, for planning and modeling complex operations on a geographical database for industrial and military flight simulators, generating reports of progress, checking integrity of world database manipulations. Improved company productivity by automating work flow processes and improving quality assurance.

Developed applications in Java and C++ for remote management of telecommunications switching hardware. Developed application using Java, JDBC, and MySQL for automating tracking of customer bug reports. Debugged product software written in C, C++, and Java.

Involved in department web site design and development. Designed and developed a bug reporting and change request database with a web front-end and a CGI Perl script back-end, which interfaced with a custom database. Designed and developed a feature request tracking system. The system was interfaced with a graphical Java applet in a web browser or stand-alone Java GUI client. The middle tier of the system was a server side Java application, which acted as a mediator between the client and the database back-end tier.