Vebjorn Ljosa Email & Phone Number

Philadelphia, Pennsylvania, Us

San Diego, Ca, Us

Extracted data from many different systems and sources (some of them GxP) and made them available to internal and external data consumers. Designed and built a unified data mesh for operational and analytical use cases (system integrations and reporting), allowing owners who understand the source domain of the data to design data products with consumers and use cases in mind.Created a process ("continuous validation") that let us build validated software with velocity. Our team consequently shipped and iterated on a high-criticality system that handled drug manufacturing QC lab results while meeting all the FDA compliance requirements.

Cambridge, Ma, Us

I rewrote BookBub’s ad selection tech, replacing a complex, slow, and inflexible Java codebase with a Clojure-based solution that brings together user interests, targeting rules, and performance estimates at run time while maintaining low latency.I also evolved BookBub’s ad marketplace by iterative analysis and A/B testing. Improved estimates of click-through rate, removed unneeded complexity and unwarranted assumptions, and in general transformed a fickle black box into a stable, well-understood equilibrium.

I rewrote Yieldbot’s ad matching tech, a low-latency, high-traffic, high-availability set of services that selects relevant ads for each pageview. Scaled the tech with the company’s explosive growth and changing needs. The code is mostly Clojure. Lots of interesting aspects: a large join that is materialized and shipped to the server clusters so that relevant ads can be found quickly; maintainable and observable logic for targeting rules; mmapable in-memory data structure for propagating click-through ratios and other statistics to be used by online machine learning; monitoring and alerting, infrastructure as code.I also optimized bidding and pricing for Yieldbot’s advertising network so as to maximize profits while meeting requirements for delivery, viewability, and targeting. Examples: linear programming to allocate expected inventory; dynamic explorations with feedback from bid acceptance rates to avoid overpaying; multi-timescale analysis to deal with volatility. Again, the code is mostly Clojure.

Cambridge, Ma, Us

Large-scale machine learning and other data-driven methods for scoring and mining complex and subtle phenotypes in image-based screens and other high-throughput experiments. Turned measurements of billions of individual cancer cells into computational profiles (“fingerprints”) that capture the biologically meaningful aspects of the cells’ appearance. Such profiles can be used to find similar genetic perturbations and/or drugs. Developed analysis methods that apply machine learning and data mining methods to large sets of complex, dirty, and poorly understood data.

Santa Barbara, Ca, Us

Developed algorithms for managing uncertainty in biological data and to obtain, search, and mine probabilistic values. Dissertation title: “Managing probabilistic data: toward data-driven biology”

Fornebu, No

Participated in development of billing software in Common Lisp.

Trondheim, No

Co-founded the company and served on the board from 1998 to 2000. Wrote user and account management software for an ISP. Planned corporate wide-area network with more than 40 locations. Set up networks and servers for ISPs and corporate customers. Provided customers with advice on network security.

Participated in development of search engine software in C++, Python and C.

Trondheim, No

Managed the central Unix system for students, with 20,000 users.

Phd, Computer Science

M.S., Computer Science