Typedef

Typedef develops Fenic, an open-source semantic DataFrame engine for developers and AI agents working with structured and unstructured data. Fenic combines familiar relational operations, such as filtering and joining tables, with model-powered extraction, classification and summarization. These AI operations have schemas and types and form part of the query plan; the execution engine handles batching, retries and response caching. Exploratory work becomes a pipeline that humans and agents can inspect, rerun and reuse. For example, an analysis of evaluation traces can run again on a new model version, with row-level lineage connecting results to their source traces.

Typedef’s work on data context for AI agents addresses a related problem in data engineering: SQL can execute successfully while producing misleading results. A join can multiply rows, and a transformation change can alter a metric’s meaning. The context-layer approach described by Yoni Michael computes table grain—the level of detail each row represents—along with lineage and relationships from transformation code. It checks join assumptions against real data and makes that context queryable before an agent acts. This gives agents information for assessing downstream effects and join semantics beyond whether a query runs.