MUMBAI, India, Aug. 12 -- Intellectual Property India has published a patent application (202641091963 A) filed by Sr University on July 29, 2026, for Adaptive Mixture-Of-Recursions Framework For Real-Time Privacy-Preserving Agentic Generative Ai-Based Dynamic Data Stream Publishing.
Inventors include Dr. K. Rajeshwar Rao; Dr. Durgesh Nandan; and Dr. S. Satyanarayana.
The application for the patent was published on August 07, 2026, under issue no. 32/2026.
Abstract: The invention provides a system comprising four integrated modules: (1) an Agentic AI Layer performing planning, reasoning, acting, memory management and on- line privacy validation over streaming windows; (2) a Mixture-of-Recursions Engine comprising a latency-aware complexity router and a bank of recursive processors (depths D0-D3) with load balancing, which predicts an optimal recursive depth for each token based on token content, context and a per-event latency budget; (3) a Streaming Privacy-Preservation Layer that continuously computes attribute/record sensitivity levels, maintains a sliding window of buffered tuples, performs incremental NSB clustering, and applies on-line bottom-up generalization to each cluster until a target (v,l)-anonymity threshold (at least v distinct sensitive values and l distinct sensitivity levels per equivalence class) is satisfied, releasing each tuple within a bounded maximum delay; and (4) a Streaming Big-Data Infrastructure layer (e.g., Apache Kafka for ingestion and Apache Flink / Apache Spark Structured Streaming for distributed, fault-tolerant processing) providing exactly-once, stateful stream processing. In operation, raw microdata tuples arriving as a continuous stream are ingested and processed on-line by the Streaming Privacy-Preservation Layer to emit an anonymized output stream with an accompanying per-window privacy certificate, each tuple being released within the configured delay bound. Agentic queries issued against this anonymized stream are handled by the Agentic AI Layer, which invokes the Mixture-of-Recursions Engine to adaptively allocate computation per token according to query complexity and the remaining latency budget, routing simple events through shallow recursion and complex events through deeper recursion so as to sustain real-time throughput. Experimental implementation on an Apache Kafka and Apache Flink / Spark Structured Streaming pipeline, over medical and financial event streams exceeding 1.5 million tuples, demonstrates: approximately 37% lower information loss compared with continuous k-anonymity and l-diversity stream baselines; approximately 42% reduction in per-event computational overhead (FLOPs) relative to static transformer baselines while improving accuracy; sustained ingestion and anonymization throughput yielding up to a 6.75x processing speedup over a traditional store-then-query RDBMS pipeline at comparable volume, with bounded (sub-second) end-to-end publication latency; and 99.2% privacy compliance using a four-level sensitivity scheme. The system is applicable to privacy-aware autonomous agents operating on data in motion in healthcare, finance and government services, where auditable, real-time privacy guarantees and bounded-latency computational efficiency are simultaneously required.
Disclaimer: Curated by HT Syndication.
