Sashank Siwakoti
Sashank Siwakoti
Senior Forward Engineer, Unit8
LLMs Inside the Data Platform: Building Production AI with Snowflake Cortex
When a compliance team at a regulated financial institution uses an LLM to auto-generate the management discussion narrative for a quarterly filing, what exactly is the audit trail? Most organizations today cannot answer that question. Data lineage tools like DBT, Apache Atlas, and MLflow trace structured data through deterministic transformations with precision but the moment a Snowflake Cortex LLM function turns governed data into narrative text, the chain of custody breaks. Which prompt template, which version, which model configuration, which parameters produced a specific sentence in a regulated report? In most environments, nobody knows. This talk presents findings from my peer-reviewed study on audit-grade prompt lineage. A gap analysis of five widely deployed governance tools shows near-zero coverage across the four most critical lineage dimensions: prompt construction capture, inference parameter logging, output hashing, and report-section mapping, a structural gap, not a configuration problem. Meanwhile, six major regulatory frameworks (EU AI Act, SOX, SEC AI disclosure guidance, Basel IV, DORA, and NIST AI RMF) now expect exactly this traceability, without specifying how to achieve it. The talk then walks through the six-layer prompt lineage framework I developed to close the gap treating prompts, inference context, model configuration, and output artifacts as first-class warehouse governance objects, with SHA-256 hash chains, Merkle root snapshots, and separation of duties enforced through native Snowflake roles. Scenario-based audit simulation shows the framework producing examiner-ready evidence across all five standard regulatory examination scenarios where existing tooling consistently fails. Attendees will leave with a deployable reference architecture and a core principle: AI-generated content in high-stakes contexts must be traceable by design, not explained after the fact.