A 500-data-mart NL2SQL talk puts orchestration—not model size—on the production agenda

Conf42 has scheduled a September 24 session whose premise is unusually specific for the text-to-SQL field: an on-premises platform said to serve more than 30 million users across more than 500 data marts with open-weight language models. Those figures are the speaker’s claims in the event abstract, not independently published measurements, and the talk has not yet premiered. That makes this a preview of evidence to watch—not a verified case study.

The architecture claim is the real signal

The abstract divides the production problem into three layers: schema retrieval and routing across hundreds of marts; SQL generation with self-correction loops; and guardrails against hallucinated queries. It also promises discussion of where agent autonomy breaks down and why larger models did not always outperform smaller alternatives.

That framing matters because it moves the deployment question away from a single accuracy score. At the scale claimed, selecting the correct data mart and supplying the relevant schema becomes a problem before SQL generation even begins. A correction loop then needs a trustworthy signal—such as parser, database or policy feedback—while guardrails must constrain what generated SQL can reach or execute. These are architectural implications of the components named in the abstract; Conf42 has not yet published the implementation details needed to evaluate them.

What practitioners should demand from the premiere

The most useful follow-up would separate the denominator behind “30M+ users” from actual NL2SQL adoption: eligible users, active users and query volume are different measures. The presentation also needs to define what counts as a data mart, describe routing-error rates, disclose the models and hardware used on premises, and distinguish syntactic self-correction from semantic correctness.

Safety evidence will matter just as much. “Guardrails against hallucinated queries” should be backed by details on read-only execution, authorization inheritance, row- and column-level controls, query-cost limits, auditability and behavior when retrieval selects the wrong schema. Without those controls, a plausible SQL string is not dependable enterprise infrastructure.

The speaker’s company site describes Vladimir Beskorovainyi as an enterprise AI architect and says his broader work has included national-scale and large-enterprise systems. It does not, on the page reviewed, identify the claimed 500-mart NL2SQL deployment or provide its evaluation data. Until the session airs, the strongest conclusion is therefore modest: the proposed agenda identifies the right production bottlenecks, while its scale and performance claims remain to be demonstrated publicly.