Most GenAI projects stall at the same spot: the model doesn't know what your data actually means. A semantic layer fixes that - a definition layer that sits on the database you already have and makes it legible to LLMs and agents. No migration. No new data team to hire.
On one production pipeline, this took text-to-SQL accuracy from ~50% to ~90%, benchmarked against a pure schema-in-context baseline.
I'm a physicist by training, an engineer by trade, and spent the last 14 years building systems and data architectures in banking and capital markets, the kind of environments where "the number is wrong" isn't an option. I'm also the creator of the open-source semantic layer semantido.
Explore my long-form articles on leveraging semantic layers, or Book a free Semantic Layer Readiness Audit
Companies I previously worked with
semantido 0.3 ships a first-class time-dimension model, alongside a complete exporter overhaul that serializes one semantic layer as OSI YAML, Markdown, and JSON. This article explains why the two belong in one release: curated metadata is the difference between an AI agent finding eight false time axes or finding the right one.
This article provides Chief Data and Analytics officers with a strategic guide for navigating the crowded semantic layer landscape, demonstrating why a robust definition system is a critical infrastructure dependency for both GenAI and agentic systems across six distinct vendor categories.
A semantic layer translates between the actual semantics of business workflows and data they produce and the LLM "understanding" gained through its training. This translation quality directly determines RAG retrieval accuracy, more so for text2SQL applications.
How I went from GenAI beginner in 2024 to launching semantido, an open-source semantic layer designed for the GenAI era.