Srivastava frames trust as an engineering property rather than a policy promise. Financial investigations can reach court years after a system produced a result, so every step must remain explainable, reproducible and auditable while sensitive consumer information stays protected.
The architecture separates data ingestion, processing and reasoning instead of treating one language model as a magic box. Kafka buffers uneven fraud traffic and keeps events in order, Spark cleans and transforms inputs, and the language model handles reasoning. Personally identifiable information is redacted before the reasoning layer, while cryptographic controls bind protected data to hardware in the server rack.
Limited offline compute makes routing essential. A semantic router sends each task to the smallest capable model, allowing routine summarization and extraction to avoid an unnecessarily large model. Srivastava says this design handled substantially more traffic without adding GPUs while reducing processing cost.
A one-way optical data diode lets new threat information enter without creating a return path for sensitive data. Incoming material first lands in quarantine for validation, and accepted records are written to an immutable, time-travel-capable store so investigators can reconstruct the state that produced an earlier decision. The talk's broader claim is that hardware boundaries, data discipline and reproducibility must be designed together for high-stakes AI.
Watch on YouTube
