Vladimir Beskorovainyi
Vladimir Beskorovainyi is an enterprise AI architect with 17 years in AI and BI, and the author of Enterprise AI: The KATA Way. He writes about applied and on-prem AI.
Askbase exists because of a measurement rather than an idea. In 2026 I published a benchmark of on-prem open-weight models on the BIRD text-to-SQL set, under a fixed protocol and with the failures included. The result that mattered was not the leaderboard position but the shape of the curve: a lightweight recipe closed part of the gap to cloud systems, and the techniques that the literature treats as essential did not survive a paired significance test.
That benchmark is also the reason this product makes a claim about accuracy at all. A number without a protocol is an advertisement.
Publications
- How Far Do On-Prem Open LLMs Get on Text-to-SQL? A Cross-Family Size x Technique Frontier on BIRD, arXiv:2606.29733. Code and results archived at Zenodo, DOI 10.5281/zenodo.20952794.
- Enterprise AI: The KATA Way, Kindle edition 12 March 2026, Amazon number one in AI. KATA is an acronym: Knowledge, Assessment, Technical Architecture, Accountability.
Work this is built on
Production NL2SQL on open-weight models, approximately 85% accuracy, over 500 data marts, on Qwen2.5-Coder-32B. That system is where the engineering questions behind Askbase came from: how a model behaves against a schema nobody designed for it, and what it does when the column names mean nothing.
A separate open benchmark, pii-roundtrip, compares anonymisation tools on frozen held-out sets with pre-registered thresholds and a run-once journal. Askbase uses the same discipline for its own measurements, and the reason is in the comparison: in this field the published numbers are mostly unverifiable.