Pie & AI: Kraków #3 - Talk to my Data

06:00 PM Opening Ceremony

06:15 PM Databricks Genie: Which Failures Tuning Actually Fixes - Aleksander Lukashou

Genie fails on messy BIRD-SQL data in two distinct ways, and only one responds to tuning. Scaffolding failures — miscounting when multiple columns could mean "customer," summing signed transaction fields without knowing negatives are credits, comparing amounts across CZK and EUR, ranking by an unintended metric — get fixed with instructions, example SQL, and trusted assets, and the same broken question starts answering correctly. The second class doesn't move: questions that need prediction, counterfactuals, or causal inference (a DoWhy problem, not a SELECT), answers that don't exist in the warehouse, and questions that were judgment calls dressed up as queries. This talk walks through the BIRD-SQL stress test that surfaced the split, what separates a fixable failure from a structural one, and why that line sits closer than most demos suggest.

07:00 PM Why your Agents need Semantic Layers, not text2sql - Konrad Bujak

Calling a table "orders" does not tells an agent whether revenue is gross, net, or after refunds. In most of the cases, text2sql fills the gap by guessing, and when it guesses wrong it returns a confident, well-formatted number instead of an error. A semantic layer defines metrics, dimensions, entities, and joins once in YAML; the agent picks a metric and an engine generates the SQL deterministically. Now a wrong pick surfaces as "I can't answer that" rather than a wrong figure on a board deck. In the 2026 dbt's benchmark Claude Sonnet 4.6 scored 90% on text2sql and 98.2% through the semantic layer, while GPT-5.3 Codex at 84.1% and 100%. But it is not all rainbows and sunshines.

07:45 AM Q&A

08:15 PM Networking

Networking is sponsored by EPAM