Intellectual Property India Publishes Patent Application for Pipebench: A System And Method For Benchmarking And Automated Optimization Of End-To-End Machine Learning Pipelines With Bottleneck Detection And Reproducibility Management

MUMBAI, India, Nov. 2 -- Intellectual Property India has published a patent application (202611109072 A) filed by Manipal University Jaipur on September 11, 2026, for Pipebench: A System And Method For Benchmarking And Automated Optimization Of End-To-End Machine Learning Pipelines With Bottleneck Detection And Reproducibility Management.

Inventors include Sanjay Agal; and Ruchika Katariya.

The application for the patent was published on October 30, 2026, under issue no. 44/2026.

Abstract: The present invention relates to a system and method for benchmarking and automated optimization of end-to-end machine learning pipelines with bottleneck detection and reproducibility management. The system comprises five interoperating software layers deployed as containerized, cloud-native components on a Kubernetes cluster: (1) a Data Ingestion & Storage layer using Apache Kafka (streaming), MinIO/Apache Hudi (versioned batch storage) and Great Expectations (data quality enforcement); (2) a Distributed Processing & Feature Engineering layer built on Apache Spark 3.4.1 with a Kubernetes operator, an online Redis feature store and an offline Hudi feature store, and a custom Spark–MLflow lineage connector; (3) an AutoML Orchestration layer using Kubeflow Pipelines 2.0 to run five hyperparameter search strategies – Random, Grid, Bayesian (HyperOpt), the novel multi-fidelity BOHB (Bayesian Optimization with HyperBand), and Evolutionary (TPOT-inspired) – each in isolated Kubernetes pods; (4) a Model Management layer using MLflow 2.8.1 with a PostgreSQL metadata backend and MinIO S3 artifact store, enforcing a staged registry with statistical validation (p 0.05) before production promotion; and (5) a Serving & Monitoring layer using TensorFlow Serving and MLflow Serving with Istio-managed canary deployments, Prometheus/Grafana observability, and online Kolmogorov-Smirnov drift detection. The layers communicate through well- defined APIs and shared storage formats (Avro, Parquet, ProtoBuf), allowing independent scaling and substitution of any component without disrupting the rest of the pipeline. The invention finds application in enterprise ML deployments, data management systems, cloud computing platforms, research institutions, and any organization that builds, deploys, or evaluates ML pipelines at scale.

Disclaimer: Curated by HT Syndication.