← All speakers

Stefano Fabbri

Conference affiliation: OpenAI

Stefano Fabbri is an applied AI engineer at OpenAI who helps financial institutions build and evaluate systems around their own data and analytical methods. Before OpenAI, he spent more than seven years at Goldman Sachs, working as a quantitative researcher across its Applied AI team and Equity ETF desk. His work spans quantitative research inside a bank and customer-facing engineering at a model developer, including API-based agents and software development with Codex.

At OpenAI, Stefano works with a team serving financial institutions in New York. That work includes helping customers define tests for a proposed system and investigate why it falls short: whether the problem lies in the model, its instructions, the information supplied to it or its connections to tools. Financial expertise matters here because an engineering test must capture what a practitioner considers a correct analysis.

Connecting models to financial practice

A joint workshop with Oliver Steffmann illustrates the setting: a wealth adviser needs to explain which holdings drove a client's portfolio gains and losses and prepare for the next meeting. The example combines internal portfolio records with financial data, reusable analytical instructions, charts and supporting citations. It connects the calculation to the explanation an adviser must review and present.

Several engineering choices in the shared session explain how that connection works:

  • Give financial data its analytical context. Access to a filing or price series is only a starting point. Tools also need to convey what a filing contains and distinguish quantities such as prices and returns, so the model selects information appropriate to the calculation. The presenters describe preparing datasets and specialized tools around those distinctions rather than simply passing raw files to a model.
  • Package methods for reuse. Skills and plugins can combine analytical instructions, data connections and examples of finished reports. A firm can supply historical spreadsheets or presentations as references, preserving its calculation and formatting conventions while adapting the analysis to a particular client. These are collaborative product and implementation efforts, rather than inventions credited to Stefano alone.
  • Make evaluation specific enough to diagnose failure. The session recommends combining financial specialists' knowledge with engineers' test datasets, then examining where a system succeeds and fails rather than relying on one aggregate score. Simple checks should precede more complicated model-based grading; otherwise, a weak evaluator can be mistaken for a weak agent. When a model does judge an output, the presenters favor explicit yes-or-no criteria over an ambiguous one-to-five rating.

These choices also affect cost. The shared session distinguishes exploratory research from recurring tasks: a more capable model may help with unfamiliar work, while a well-specified skill can make a less expensive model suitable for repeated analysis. Evaluation determines whether that substitution preserves the quality the task requires.

1 conference talk

Loading more talks and highlights…

References