Guardians of the State: An Air-Gapped AI Fortress for Consumer Data — Rachna Srivastava, DFPI
California’s financial regulator built an offline AI pipeline around replayable data, hardware-backed redaction, model routing, and one-way updates for court-defensible evidence.
DFPI separated ingestion, processing, and reasoning across Kafka, Spark, and local models after an initial system failed within **2 hours**. Routing work to the smallest capable model handled **over 80% of tasks** and tripled traffic without more GPUs.
For sensitive agent systems, make provenance and replay architectural requirements. Clean inputs before inference, tokenize PII with hardware-held keys, route by task, quarantine external updates, and preserve the exact data state behind each decision.
DFPI separated ingestion, processing, and reasoning across Kafka, Spark, and local models after an initial system failed within **2 hours**. Routing work to the smallest capable model handled **over 80% of tasks** and tripled traffic without more GPUs. For sensitive agent systems, make provenance and replay architectural requirements. Clean inputs before inference, tokenize PII with hardware-held keys, route by task, quarantine external updates, and preserve the exact data state behind each decision. The physical isolation, one-way data diode, and immutable history target unusually high legal and privacy stakes. The talk reports nearly **70% lower processing cost**, but provides no workload definition or independent security and accuracy evaluation.
This turns the candidates’ general requirements for governed, sandboxed enterprise agents into a high-stakes architecture with separated stages, one-way transfer, tokenized PII, replayable state, and local model routing. It confirms that provenance and enforcement belong below the model, while narrowing the reported efficiency gains to an unspecified workload without independent security or accuracy validation.