International Journal on Science and Technology

E-ISSN: 2229-7677   •   Impact Factor: 9.88

A Widely Indexed Open Access Peer Reviewed Multidisciplinary Bi-monthly Scholarly International Journal

Call for Paper Volume 17 Issue 2 April-June 2026 Submit your research before last 3 days of June to publish your research paper in the issue of April-June.
Submit Research Paper Join as a Reviewer

Plagiarism is checked by the leading plagiarism checker TurnItIn

Call for Paper

Volume 17 Issue 2
April-June 2026

Submit your research paper

Indexing Partners
Academia Advanced Sciences Index Bielefeld Academic Search Engine CiteSeer DRJI Google Scholar Independent Search Engine & Directory Network ISI (International Scientific Indexing) Issuu Mendeley Research Networks
RefSeek ResearcherId - Thomson Reuters ResearchGate Scirus Scribd Semantic Scholar UTeM - Universiti Teknikal Malaysia Melaka Wiki for Call for Papers WorldCat Zenodo

KAIZEN: Governed Continual Improvement for LLM-Backed Enterprise Systems Drift Detection, Intervention Selection, and Progressive Delivery for Multi-Artifact LLM Applications

Author(s) Sandeep Nutakki
Country United States
Abstract Large language model (LLM) applications change after launch through prompts, retrieval indexes, tool schemas, routing rules, guardrails, and fine-tuned weights, yet many MLOps practices still monitor only checkpoints and aggregate accuracy. This paper presents KAIZEN, an architecture and controlled replay benchmark for governed continual improvement of LLM-backed enterprise systems. KAIZEN combines behavioral-semantic drift detection, budget-aware intervention selection, human-in-the-loop curation, and risk-tiered progressive delivery. We evaluate KAIZEN in a controlled 18-month replay spanning 43,200 synthetic enterprise requests, 36 injected drift events, 54 candidate releases, and 1,440 adjudicated evaluation cases. The study is not a production deployment; it is a reproducible replay under known ground truth. In the controlled replay, KAIZEN improved incident detection F1 from 0.58 to 0.82 under the stated workload and drift assumptions, and detected degradation 5.7 days earlier on median (95% clustered bootstrap CI: 4.3-7.1, p < 0.001). Under the same assumptions, KAIZEN reduced unnecessary retraining actions by 38.9% relative to scheduled monthly retraining (95% CI: 29.4-47.6, p = 0.002) and lowered total simulated improvement cost by 46.2% (95% CI: 39.4-52.8, p < 0.001). Compared with direct rollout in the replay, progressive delivery reduced simulated user-visible regression exposure from 19.6% to 3.1% of affected traffic (95% CI for relative reduction: 79.1-88.4, p < 0.001). These results support KAIZEN as an architecture and benchmark design; they do not establish field performance.
Keywords MLOps, LLMOps, model monitoring, concept drift, large language models, continual learning, progressive delivery, human-in-the-loop learning, model governance
Field Engineering
Published In Volume 17, Issue 2, April-June 2026
Published On 2026-06-04
DOI https://doi.org/10.71097/IJSAT.v17.i2.11319
View / Download PDF File

Share this


CrossRef

CrossRef DOI is assigned to each research paper published in our journal.

IJSAT DOI prefix is
10.71097/IJSAT

Downloads
Research Paper Format Copyright Permission Form and Undertaking Form Cover Page Vol 17 Isu 2Cover Page Vol 17 Isu 1Cover Page Vol 16 Isu 4

All research papers published on this website are licensed under Creative Commons Attribution-ShareAlike 4.0 International License, and all rights belong to their respective authors/researchers.

CC-BY-SA Open Access

About IJSAT
Fees & Payment
Current Issue
Publication Archive 		Submit Research Paper
Track Submission Status
Publication Guidelines
Publication Ethics
Peer Review & Plagiarism 		Join as a Reviewer
Editors & Reviewers
Reviewer Referral Program
Get Reviewer Membership Certi. 		Website/Journal Policies
Usage Policy
Content Policies
Privacy Policy
Contact Us Message on WhatsApp +91-9687-182-185 editor@ijsat.org