Multi-Layer AI Architecture for Real-Time Detection of Workflow and Data Integrity Failures in Distributed Retail Platforms
Type: article
Abstract: Distributed retail systems run on heterogeneous data systems whose joint state represents business invariants that cannot be verified in the life-cycle of any one of them.Silent inconsistencies arise when, downstream, purchase orders go missing, quantities do not match, suppliers are blocked, or data drifts across systems.Rule-based monitoring systems and singlesystem observability do not detect these failures by definition.Text-to-SQL and retrieval-augmented generation technologies answer these questions without executing the multi-step investigation an incident demands.We present a two-layer AI architecture that separates deterministic data access and nondeterministic reasoning.The semantic data access layer, implemented as an MCP server, defines eight domain tools as well as five templates with natural-language descriptions.The stateful reasoning layer parses incidents, reasons about them, selects tools, takes investigative steps and derives remediation following a seven-step workflow captured as a knowledge-based process in a Camunda BPMN process engine.Production evidence from a large international retailer shows that average resolution effort per incident fell from 9 hours to 2.25 hours, that Level-1 incident volume fell from 33 to about 8 over comparable 2-month windows, and that engineering escalations fell from 67 to about 17 over comparable 3-month windows, a reduction close to 75% in both effort and volume.Tool description debt is a new class of technical debt that is specific to LLM-based observability systems.A four-element taxonomy is proposed.
Source: CyberLeninK (CyberLeninka)
Author: Sai Sruthi Puchakayala
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223496528