About

Currently I focus on building Spark and other AI solutions, mostly for deep tech applications.

Company building

I build Spark, an AI-based solution for technology discovery. Currently used at Chevron, Merck, Siemens, and others.

Founder and CEO of Mergeflow. Initially built as a project business. Then transformed it into a 100% SaaS product business, from $0 to mid-six-figure ARR, using only revenue-based funding.

At Mergeflow, acquired multi-year SaaS contracts from customers like BASF, Bayer, Beiersdorf, BMW, Covestro, Dell Technologies, Leica, LG Chem, Siemens, thyssenkrupp.

Built an AI-based business idea generator. Performance is comparable with human consultants. Customer: Siemens Advanced Manufacturing.

For Mergeflow, secured a six-figure EUR amount of state and federal government R&D funding.

Machine learning and natural language processing

Developed a stateless ML system at Mergeflow for processing diverse documents (webpages, news, patents, scientific publications) and extracting key topics and entities.

Developed a TRL estimation method using web data analysis with University of Toronto's Ready Lab.

At the University of Cambridge, built a tool for annotating biomedical texts. Used the data to develop and train ML algorithms for extracting gene names from texts, and for connecting them to microarray data.

At MIT, in collaboration with Lincoln Laboratory, developed new methods for measuring, and guiding further improvement, of the quality of speech-to-text transcripts and machine translation. Funded by DARPA (EARS and TIDES programs).

Algorithm design and optimization

Developed a fast entity matching algorithm for vendor addresses. Reduced runtime from one day to 40 seconds (customer: Logitech).

At Mergeflow, my team and I redesigned an algorithm for extracting chemical compound names from texts. This redesign reduced monthly cloud computing costs from $20,000 to $50.

Finance analytics

Developed a new method for measuring hedge fund performance persistence, based on runs and scans statistics.

Built a system to extract and standardize hedge fund data (returns, AuM, etc.) from various sources like emails, documents, and spreadsheets.

Ran analyses comparing human vs. algorithmic hedge fund performance ratings. The result was an optimized combination of human and algorithmic ratings, resulting in a substantially improved overall hedge fund due diligence process.

Customer for the above: UniCredit.

Cybersecurity and intelligence analysis

Built an OSINT system for cyber security analysis with the University of the Armed Forces (Munich, Germany), Airbus Defence and Space, and various federal agencies. The system collected and analyzed data from various open sources.

Education

PhD in Brain and Cognitive Sciences at MIT (funded by DARPA).

Postdoc at Cambridge University (Computer Laboratory and Genetics Department).

Publications

Together with Edward Gibson, wrote a book on data structures for representing coherence in natural language, published by MIT Press.

Published papers, talks, and posters at EACL, ACL, in Computational Linguistics, and in MIT AI Technical Reports.