Illustrative engagement across five service pillars and five industries, reflecting representative Meruksha engagement patterns.
Industry: Industrial Manufacturing
Client Profile
The client is a mid-size industrial manufacturer operating multiple plant sites, with an engineering archive spanning more than 30 years of process documentation, equipment manuals, maintenance logs, and incident reports. The company’s OT/ICS environment is subject to strict network segmentation policies overseen by a dedicated plant security function.
Challenge
Process engineers across the company’s plant network wanted an AI system capable of answering natural-language questions against three decades of accumulated internal documentation — process specifications, equipment manuals, past incident investigations, and maintenance histories — much of which existed only as scanned documents or informal notes with no consistent indexing. The obstacle was not technical feasibility but data sensitivity: this archive contained proprietary process engineering trade secrets and, in some cases, configuration details tied to the plants’ industrial control systems. Any AI approach built on a public API-based large language model was immediately ruled out by the security team, since it would require transmitting this sensitive material outside the corporate network to a third-party provider. Two prior attempts at an AI documentation assistant, both proposed by outside vendors using commercial cloud AI APIs, had been rejected outright by the CISO and the plant OT security lead before reaching a pilot stage, leaving engineers without any AI assistance and continuing to search decades of paper and PDF records manually.
Approach
Meruksha proposed and deployed a self-hosted, open-source large language model — a fine-tuned variant of Llama — entirely within the company’s own virtual private cloud, ensuring no engineering documentation or query data ever left the corporate network boundary. A retrieval layer was built directly over the digitized and newly OCR-processed engineering archive, with all embeddings generated on-premises rather than through any external embedding API. Recognizing the OT security team’s specific concern about configuration data, Meruksha configured explicit guardrails preventing the model from surfacing detailed industrial control system network configuration information even to authenticated internal users, mirroring the same network segmentation logic already enforced on the plant floor. The fine-tuning process incorporated the company’s own historical incident reports and engineering terminology, so the model’s responses reflected the organization’s specific equipment and process vocabulary rather than generic industry knowledge.
Outcome
Engineers gained natural-language search capability across more than 30 years of process documentation with a verified zero-data-egress architecture, satisfying both the CISO and the plant OT security team — the two stakeholders whose objections had killed both prior AI initiatives before they reached pilot stage. Mean time to locate relevant historical incident data for a given equipment failure dropped from what engineers described as multi-day manual archive searches to a matter of minutes. The security-first architecture Meruksha implemented has since been referenced internally as the template for evaluating future AI tools proposed for use in or near the company’s OT environment.