MUMBAI, India, Oct. 12 -- Intellectual Property India has published a patent application (202621088300 A) filed by Persistent Systems Limited on July 20, 2026, for System And Method For A Historical Learning And Continuous Optimization For Ai Workflow Execution.
Inventors include Mr. Nitish Shrivastava; and Mr. Pradeep Kumar Sharma.
The application for the patent was published on October 09, 2026, under issue no. 41/2026.
Abstract: A historical learning and continuous optimization system (100) converts agentic-workflow execution traces into a versioned historical execution memory in which each decision record links task features, workflow and node versions, candidate actions with selection probabilities, selected action and context, resource usage, verification results, human labels with correction magnitude, downstream outcomes, and provenance. A point-in-time feature store (140) materializes features as known at each decision, and immediate and delayed labels are retained with source and reliability across a five-level label hierarchy. Stage- and action-specific predictors estimate cost, quality, retry, escalation, verifier-rate, branch, and acceptance outcomes with uncertainty. An off-policy evaluator (180) estimates a proposed policy’s value using propensity, inverse propensity weighting, doubly robust, and matched-replay methods before deployment. A champion deployer (410) and challenger deployer (420) expose policies under bounded exploration and guardrails (440), and a drift monitor triggers rollback (470), while privacy and tenant partitions enforce isolation, retention, and deletion.
Disclaimer: Curated by HT Syndication.
