TriageGPT for DataOps: Large Language Models as Intelligent Incident Classification and Resolution Agents in Data Pipeline Operations

Purpose: This paper presents TriageGPT for DataOps, an architecture in which large language model agents classify data pipeline incidents, retrieve relevant context, and perform constrained autonomous remediation, in order to reduce the manual triage effort that drives incident duration and on-call fatigue. Methodology: TriageGPT combines a retrieval-augmented classification agent that assigns alerts to a data pipeline specific incident taxonomy, a tool-use resolution agent that invokes actions from a vetted, blast-radius-tiered action library, and an audit-logged guardrail layer that gates and can automatically roll back actions, paired with confidence-based human-in-the-loop escalation. We describe the architecture, algorithms and calibration policy, map the design to safety and governance guidance, and report a prototype evaluation on a benchmark pipeline with a fourteen-category taxonomy, 2,800 labeled synthetic incidents and a twelve-week comparative deployment trial. Findings: TriageGPT achieved 94.3 percent top-one classification accuracy and 95.1 percent factual attribution accuracy with retrieval, against 68.4 percent without it. It resolved 71.5 percent of test incidents without human intervention at an incorrect-action rate of 0.6 percent, every one of which the guardrail caught and reversed. In the deployment trial it reduced median time to resolution from 41 minutes under manual on-call practice to 9.8 minutes. Unique Contribution to Theory, Policy and Practice: For theory, the paper treats autonomous remediation as something earned through calibrated confidence and bounded blast radius rather than assumed. For practice, it provides an architecture, algorithms and evaluation method for resolving routine pipeline incidents while deferring the rest to a human with an evidence-grounded case. For policy, it maps autonomous remediation to existing guidance, informing how organizations set guardrails, audit requirements and risk tolerance for AIOps agents.

Authors

Publication Details

Journal
International Journal of Computing and Engineering
Published
2026-11-19
DOI
https://doi.org/10.47941/ijce.4117
Primary Topic
Software System Performance and Reliability
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

TriageGPT for DataOps: Large Language Models as Intelligent Incident Classification and Resolution Agents in Data Pipeline Operations

Prabhu Ravi, Rajesh Thondamanati
International Journal of Computing and Engineering
Software System Performance and Reliability
article

TriageGPT for DataOps: Large Language Models as Intelligent Incident Classification and Resolution Agents in Data Pipeline Operations

Prabhu Ravi, Rajesh Thondamanati
article en

Abstract

Purpose: This paper presents TriageGPT for DataOps, an architecture in which large language model agents classify data pipeline incidents, retrieve relevant context, and perform constrained autonomous remediation, in order to reduce the manual triage effort that drives incident duration and on-call fatigue. Methodology: TriageGPT combines a retrieval-augmented classification agent that assigns alerts to a data pipeline specific incident taxonomy, a tool-use resolution agent that invokes actions from a vetted, blast-radius-tiered action library, and an audit-logged guardrail layer that gates and can automatically roll back actions, paired with confidence-based human-in-the-loop escalation. We describe the architecture, algorithms and calibration policy, map the design to safety and governance guidance, and report a prototype evaluation on a benchmark pipeline with a fourteen-category taxonomy, 2,800 labeled synthetic incidents and a twelve-week comparative deployment trial. Findings: TriageGPT achieved 94.3 percent top-one classification accuracy and 95.1 percent factual attribution accuracy with retrieval, against 68.4 percent without it. It resolved 71.5 percent of test incidents without human intervention at an incorrect-action rate of 0.6 percent, every one of which the guardrail caught and reversed. In the deployment trial it reduced median time to resolution from 41 minutes under manual on-call practice to 9.8 minutes. Unique Contribution to Theory, Policy and Practice: For theory, the paper treats autonomous remediation as something earned through calibrated confidence and bounded blast radius rather than assumed. For practice, it provides an architecture, algorithms and evaluation method for resolving routine pipeline incidents while deferring the rest to a human with an evidence-grounded case. For policy, it maps autonomous remediation to existing guidance, informing how organizations set guardrails, audit requirements and risk tolerance for AIOps agents.

International Journal of Computing and EngineeringVol. 8(4)
Software System Performance and Reliability
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon