Jon Chun: AI evaluation researcher & builder
Jon Chun is a researcher and builder working at the frontier of AI evaluation. His current work examines model reasoning, human judgment, agent behavior, explanation, and evaluation under ambiguity.
His career combines empirical model evaluation, computational methods, deployed technical systems, research infrastructure, and interdisciplinary programs that make people capable of conducting original AI research.
Frontier AI evaluation
Jon serves as Co-PI representing the Modern Language Association in the NIST CAISI consortium. His standards-facing work includes model evaluation, reasoning and judgment, explanation, agent behavior, and red-teaming.
He is also Co-Principal Investigator of Archival Intelligence, part of the Schmidt Sciences Humanities and AI Virtual Institute. His work includes computational workflows, evaluation design, and technical research infrastructure for AI systems operating over difficult multilingual and historical materials.
These projects bring earlier habits into current evaluation research: define the construct precisely, preserve the evidence behind a result, test behavior under ambiguity, and treat failure modes as part of system design.
Building systems and research infrastructure
Jon’s technical background includes both deployed production systems and current research infrastructure.
He studied electrical engineering and computer science at UC Berkeley EECS, co-founded SafeWeb in 2000, and later served as CEO. SafeWeb developed privacy and security infrastructure and was acquired by Symantec in 2003. Jon continued technical development at Symantec and holds two US patents related to browser-based SSL and clientless VPN systems.
At Archival Intelligence, he now works on computational workflows, evaluation design, and technical infrastructure for difficult source materials. These experiences inform his current approach to AI evaluation: systems are judged by how they behave with real users, uncertain inputs, operational constraints, and changing environments.
Building methods
Jon’s independent research has developed computational methods for measuring complex behavior across long-form and multimodal data.
- SentimentArcs. An independently authored ensemble method for narrative sentiment analysis and trajectory comparison.
- Narrative trajectory comparison. LTTB normalization and dynamic time warping make unequal-length trajectories computationally comparable.
- Explainable narrative analysis. Joint work with Katherine Elkins adds sentence-level evidence and interpretable comparison.
- MultiSentimentArcs. An independently authored multimodal method comparing emotional trajectories recovered from dialogue and images.
Explore the full research program → · See how the methods have been used →
Expanding who gets to shape technology
Jon has also spent much of his career building technical capability in other people and interdisciplinary research teams.
At Kenyon, he co-founded Human-Centered AI and designed the technical curriculum architecture that enabled students without traditional computer-science backgrounds to conduct original computational and AI research.
The program combines programming, data analysis, model evaluation, domain expertise, and research practice in repeatable technical workflows. Students progress from first computational projects toward research competence: framing questions, building systems, testing behavior, interpreting evidence, collaborating across disciplines, and publishing original work. See the curriculum and mentored research →
Selected facts
- NIST CAISICo-PI, MLA team, NIST CAISI consortium
- Schmidt SciencesCo-PI, Humanities and AI Virtual Institute work on Archival Intelligence
- Kenyon Collegevisiting instructor of humanities and affiliated scholar in scientific computing
- EducationB.S., Electrical Engineering and Computer Sciences, UC Berkeley; M.S., University of Texas at Austin
Profiles: Google Scholar · ORCID · GitHub · LinkedIn · CV