Skip to main content
bash TV

Guardians of the State: An Air-Gapped AI Fortress for Consumer Data — Rachna Srivastava, DFPI

AI Engineer

1.5K views29 Aug 2026

YouTube

The fiber optic cable carrying data into California's financial fraud system has been cut in half. One end sits on the internet with a laser transmitter. The other end, inside the building, has only a receiver. There is no transmitter pointing outward, so data physically cannot leave. Rachna Srivastava's team chose that over a software firewall for a blunt reason: any configuration can be misconfigured, and a misconfigured secure system is an exploited one. Everything her group builds has to survive a defense attorney whose entire job is attacking it, which means every step must be explainable, reproducible and auditable years later. They did not get there gracefully. The first build was the obvious one, an open model in an isolated environment with guardrails, and it collapsed within two hours. The diagnosis was not a weak model. They had treated it as a magic box rather than a data pipeline, so Kafka took ingestion and replay, Spark took cleaning, and the model was left to reason over data that had already been made sane. Her framing is that most AI data problems are data engineering problems wearing an AI mask. Running one frontier model for every task was making a neurosurgeon take everyone's blood pressure, so a router now sends over 80% of work to the smallest model that can do it, tripling throughput on the same GPUs. Speaker info: - https://www.linkedin.com/in/rachana-srivastava-ms-mba-78bab86 - https://dfpi.ca.gov/ Timestamps: 0:00 - When seeing and hearing stopped being proof 2:32 - Building for the courtroom, not the demo 3:44 - Why encryption and private endpoints were not enough 6:08 - Certifications as paper 7:22 - The first build, and two hours to collapse 8:29 - Kafka for spikes, ordering and replay 10:56 - Data engineering problems wearing an AI mask 12:01 - A key bolted to the server rack 13:13 - One frontier model doing every job 14:23 - Routing 80% of work to smaller models 15:38 - Learning without opening a hole 16:53 - A cable cut in half 18:05 - Time travelling to the moment of a decision 20:27 - Trust as a physical property

Join the discussion

Sign in to join the discussion

Sign in