▶ Watch ↗19:34
Conference affiliation: Nori
Clifford Ressel is co-founder and CTO of Nori Agentic, where he builds infrastructure for teams running fleets of agents in the cloud. His work connects agent environments with the data problems he encountered in back-office finance: reconciling inconsistent records and determining whether a proposed change has enough evidence behind it.
Before Nori, Cliff worked on the data platform at the hedge fund Freestone Grove as a Member of Technical Staff and held a Staff Software Engineer role at Roche. His public profile connects that background with his work at Nori and his technical writing.
Nori gives agents cloud environments with repositories, dependencies, tools and services ready to use. Integrations and triggers let work begin without someone opening a local terminal; the result can be a pull request, a document or another completed task for review. The infrastructure supports different coding agents rather than requiring teams to adopt a single harness.
Two collaborative company projects address the practical differences between those agents:
Cliff’s individual open-source contributions include a correction to Codex’s stream-error handling, merged in November 2025.
In his March 2026 essay on implementation and verification costs, Cliff distinguishes producing code from establishing that it behaves correctly. He describes changes in his own workflow and advocates investing agent-assisted productivity gains in behavioral testing and observability. Cheaper implementation creates room for more examination of the resulting behavior.
His explanation of investment data makes the verification problem concrete. In a joint presentation with Nori co-founder and CEO Amol Kapoor, Cliff describes the security master—the central database of an investment firm’s securities and instruments—as an organizational primitive between data producers and consumers. Conflicting values and mapping errors should be resolved there before they spread into risk calculations, profit-and-loss accounting and regulatory reporting.
Google’s publicly traded shares illustrate why this requires more than matching identifiers. Information about the company, a share class and an exchange listing belongs at different levels. A dataset describing the legal entity cannot simply be joined to trading records indexed by a security identifier. Preserving the relationships between those levels lets analysts connect records without confusing what each one describes.
Private assets introduce a harder problem. A loan financing a power plant may have no identifier shared by everyone involved. Its repayment schedule and collateral live in individual agreements and amendments; the documentation defines the asset. Each new investment therefore brings its own body of evidence to reconcile, rather than another instance of a standardized instrument.
Amol separately described the Nori team’s work with an asset manager, where an agent gathered supporting material, cited its findings and recommended whether to retain a record change. A human record owner reviewed the recommendation. Cliff’s contribution to the explanation was the underlying data model: before evaluating a changed value, the system must establish which asset or entity it describes and how the source maps into the firm’s records.
▶ Watch ↗19:34
Loading more talks and highlights…