Summary
Rachna Srivastava, a technology leader at California’s Department of Financial Protection and Innovation (DFPI), presents the architecture behind an air-gapped AI system built to investigate financial fraud affecting 39 million Californians. The talk centers on a fundamental constraint: every piece of evidence the system produces must be reproducible and defensible in court, which rules out cloud-based deployments and conventional encryption approaches.
Srivastava walks through the hard lessons learned when the team’s initial attempt — a simple offline model with system prompts — collapsed within two hours. The fix required treating AI as one component in a proper data pipeline: Apache Kafka handles bursty ingestion and provides the event-replay capability that serves as courtroom evidence, Apache Spark cleans heterogeneous inputs (bank statements, audio files, screenshots) before they reach the model, and an LLM handles only the reasoning layer. PII is protected using a SHA-256 cryptographic vault with a hardware security module physically attached to the server rack.
Two memorable lessons emerge: most AI data problems are actually data engineering problems in disguise, and when stakes are high, trust hardware over software. The talk is a rare look at how a government agency is navigating air-gapped GPU capacity planning, reproducibility requirements, and the evidentiary standards that private-sector AI deployments rarely face.
📺 Source: AI Engineer · Published August 29, 2026
🏷️ Format: Deep Dive
