Helping advance the state of the art of retail AI, at the world's largest retailer.

slowest rolling "stealth startup" you ever did see!(ie mostly consulting)

Advising early stage (pre-seed/seed) tech startups in the AI/ML space

Helping shape the future of Machine Learning Infrastructure on which all of LinkedIn's AI-based applications and features are built.Initially responsible for the technical vision and architecture for the infrastructure-side of a company-wide (~dozen teams across 3 top-level engineering orgs) project to enable "embedding-based retrieval" (vector similarity - aka "Vector DB") to power a wide variety of recommender systems, ad-delivery systems, search products, and of course: short + long term external memory for LLMs in GAI use cases.Later, when we spun up a new tech organization merging model serving, feed recommendation infra, and search, jumped into building the next generation of our model inferencing, information retrieval, and recommender systems stack, enabling things like multi-engine and model-based retrieval, RAG for agentic workflows, LLMs-as-ranking models, and everything in between.

LLM Crash Course: Introduction to LoRA and Agentic ApplicationsIntroduced ML, Deep Learning, and transformers to CS undergraduates, focusing on implementation and practical use with PyTorch, including fine-tuning Phi2, building simple RAG pipelines, and working with LangChain.

"The Software Industry: A Crash Course"Teaching graduating seniors the fun joys of git, code reviews, TDD (or not), CI/CD, Service Oriented Architecture, micro-services vs macro-services, and all the things they don't teach you in school. Just in time to terrify them before going out into the industry :)

Responsible for the long-range vision of AI and Search at Salesforce.Mentoring, helping hire, guide the Virtual Architecture Teams toward building the right Search and AI infrastructure where intelligent data-driven applications are powerful, scalable, and easy to build, regardless of whether it's done by a specialized internal team of ML Engineers and Data Scientists, external partner ISVs, or via low-code solutions for our customer admins.

As the Search Relevance organization expanded to include the Einstein Language Intelligence and Einstein Platform Apps teams, my role expanded with it, to design and lead implementation of generic recommender and search applications in composable ways, incorporating more deep learning-based models for language-understanding, text embedding, and ranking.Worked on major initiatives toward incorporating the following pieces of functionality into our systems:* Vector Search as a Service ( <-- HNSW graphs via NMSLib in OpenSearch)* Bring Your Own Model ( <-- via NVidia Triton and others coming soon)* Inference Graph Execution Service ( <-- home-grown )* Feature Store ( <-- via Feast + HBase as a backend)

Lead the technical vision for the Search Relevance organization, including the following teams:* Search Ranking* Search Query Understanding* Search Data and Analytics* Search Intelligence PlatformLots of fun projects building up our model training platform, migrating from simple linear models stored as JSON blobs in Oracle tables to sophisticated deep learning models leveraging personalization and tenant clustering, including kicking off our deep learning library for search related problems (such as query classification and document reranking): https://github.com/salesforce/ml4ir

Head up company-wide efforts around scalable data processing and applied machine learning infrastructure in the context of search and related data-driven products, bridging the gap between R&D and product engineering. My primary focus was bringing the "intelligence" to intelligent semantic search for the enterprise: get the right actionable enterprise data available (for the right context) to the end user, utilizing deep understanding of the content, the user's interaction history (personalization), and the current search query session, tied together with a set of query understanding modules, personalization stores, and learning to rank models.Partially outward focused - evangelism, training and advising (both internal and external), as well as developing prototypes for customers pushing the edge of the complexity envelope; and partly inward focused: once a product-modification/extension has been seen to be useful in N > 2 important customers, shepherd the update fully into the product suite, after deeper testing, hardening, and scale-proofing.Team areas of interest: applied machine learning, recommenders, personalization, relevance tuning, data quality analysis, and query/content analytics, NLP; deep learning as applied to all of the previousTechnologies frequently interacted with: Spark (+SparkML), Solr (+Lucene), AWS, kubernetes, keras/tensorflow, sklearn, scala, java, python.

Search, relevance, ranking, applied NLP and machine learning to search and recommender products on top of an open source stack (Lucene/Solr, ZooKeeper, Spark/MLLib, Mahout, etc).

Data infrastructure engineering and applied machine learning to help scale out applying NLP techniques on semantic search over a corpus of scientific articles, launched in the fall of 2015: http://www.semanticscholar.org/Scala, Akka, Spark, ElasticSearch, etc.

Multiple roles, variety of teams. Fun stuff. See specific role positions below for more details.Tech Lead, User Modeling:Built and lead a team of distributed-systems and machine learning engineers responsible for understanding Twitter's "interest graph", taking tweet text, links, and the social graph, and using a variety of classification and topic modeling techniques to build better personalized relevance for various products at the company.Integrated topic-based personalization systems into the #discover page at Twitter using a mixture of (SGD / logistic-regression based) text classification, graph label propagation, and learning to rank on implicit user feedback engagement/impression data (see MLConf talk here[1]).Tech Lead, User Search:Designed and built the self-healing distributed search system for finding user accounts relevant to name / topical queries on twitter.com and the twitter API. Technology we used to build this includes: lucene, hadoop, pig, zookeeper and thrift, and much of it is already open-sourced[2], with more to come at some point. Co-authored a paper [3] on the distributed systems side of this work. Lead the team responsible for the maintenance and improvement of this aspect of Search at Twitter.----[1]http://www.slideshare.net/SessionsEvents/jake-mannix-m-lconf-2013[2]https://github.com/twitter/commons[3]http://www.umiacs.umd.edu/~jimmylin/publications/Leibert_etal_SoCC2011.pdf

Served as Chair of Mahout's Project Management Committee, helping make sure that ASF processes are followed by the large and varied Mahout community, involving not only technical advice and help for new users (as all contributors do), but advising users regarding regarding proper use of Apache copyrights and trademarks, and advocating for Mahout adoption within my "day job" at Twitter.Looking for open source scalable machine learning on your vanilla Hadoop cluster? For more details, visit us at http://mahout.apache.org

The Apache Giraph project is a fault-tolerant in-memory distributed graph processing system which runs on top of a standard Hadoop installation, and is capable of running any standard Bulk Synchronous Parallel (BSP) operation over any large generic data set which can be represented as a graph, and is a loose implementation of Google's Pregel.

Helping build an open-source, commercial friendly licensed, scalable machine learning library. Have focused on improving our bayesian topic modeling work (e.g. LDA) and linear algebra primitives, as well as adding dimensional reduction components for NLP and recommender systems.

One of the primary architects of the (original - long since replaced, several times over!) distributed, real-time, faceted people search platform you most likely used to find this profile.Helped justify, scope out, and build a new engineering team inside of LinkedIn's analytics-engineering organization: from two initial engineers, we built a Recommendation Engine team of roughly (depending on how you look at the org chart, at the time I left) ten engineers, data scientists, and managers.While interviewing and hiring for this team, created the infrastructure for a generalized entity-to-entity realtime (i.e. results computed online) recommendation system, using a variety of content and usage-based matching techniques with machine learning models trained on our Hadoop cluster.Implemented and launched a handful of recommendation-based products on top of this infrastructure, including Talent Match ("People for your Job posting"), Jobs You Might Be Interested In, and helped guide into production a half-dozen more as we scaled up the team.I spent some of my work-time on a couple of extension projects built on top of Apache Lucene:* Core committer on the high-performance open-source faceted search library, BoboBrowse (http://bobo-browse.googlecode.com).* Committer on the open-source real-time search and indexing system Zoie (http://zoie.googlecode.com).Creator/maintainer of the nascent open-source NLP / graph-theoretic matrix library Decomposer (http://decomposer.googlecode.com) (since absorbed into Apache Mahout)[note: in 2008, the 4 grades of eng at LinkedIn were "SDE", "Sr. SDE", "Principal SDE", and "Architect". That's it. I interviewed and was told I was "on the line between Sr. SDE and Principal, and made my case to get Principal, and got it. In 2012, LinkedIn re-graded and the better estimate of what my level would have been (after I left!), is "the level above Sr SDE" which today is called Staff SDE]

Directed architectural vision for all things Search-related, from redesigning, developing, and maintaining a high-availability query-expanding job-post (full-text) search engine (as a replicated Tomcat+Lucene+Spring+mySQL-based web service), to R&D of next-generation conceptual applicant/position matching technology, using a ngram-based partially-parallelized (single-box, not ready for Hadoop without algorithm modification) eigen-decomposition algorithm and user feedback for assisted machine learning to provide a personalized search experience.Designed and implemented a JRuby-based Rails plugin as glue to transparently allow vanilla-seeming ActiveRecord models to say they "acts_as_conceptual" while being coded into simple RoR apps which can scale and perform as J2EE apps.Lead and mentored junior developers along their technical career path, and participated in business-space technical decision making (buy/build/partner) with CxO-level management.

Performed theoretical high energy physics and cosmology research for Stanford University's Institute for Theoretical Physics: developed and tested inflationary cosmology simulation software, with C++/optimized ASM for the computational back-end, Hibernate / MySQL datastore, and Jakarta Struts MVC (Servlet + JSP) for front-end web presentation. Computed numerical differential equation integration (and analytic approximations) for cutting-edge theoretical dark energy models.

RealServer streaming audio/video development in C++

Client and Server streaming audio/video streaming applications

Member of the inaugural class (2 students picked each year) of visiting graduate fellows at the newly formed Kavli Institute for Theoretical Physics (coincident with the Strings '99 Conference), working with Philip Argyres and his student on string theoretic representations of 4d SCFTs which allowed for a geometric proof of electric-magnetic duality in systems less than maximal supersymmetry.