[Paper] A Human-Centred Architecture for Large Language Models-Cognitive Assistants in Manufacturing within Quality Management Systems

Source: arXiv - 2603.16325v1

Overview

The paper proposes a human‑centred software architecture that lets manufacturers plug Large Language Model‑based Cognitive Assistants (LLM‑CAs) into their Quality Management Systems (QMS). By focusing on modularity, scalability, and user‑feedback loops, the authors aim to turn cutting‑edge LLM capabilities into practical, day‑to‑day tools for continuous improvement on the shop floor.

Key Contributions

• A component‑based reference architecture tailored to QMS workflows, explicitly designed for LLM‑CA integration.
• Requirement‑driven design process that captures both functional (e.g., data ingestion, traceability) and non‑functional (e.g., security, explainability) needs of manufacturing stakeholders.
• Iterative validation through expert focus groups, ensuring the architecture aligns with real‑world quality engineers, line operators, and IT staff.
• Guidelines for flexibility and work augmentation, showing how the architecture can evolve with new LLM models or changing regulatory requirements.
• Roadmap for industrial pilots, outlining steps to move from prototype to deployment with partner factories.

Methodology

1. Requirement Analysis – The team conducted workshops with quality managers, engineers, and IT personnel to extract concrete needs (e.g., audit‑trail generation, multilingual support, real‑time feedback).
2. Architecture Design – Using a component‑based approach, they defined four layers:
   • Data Acquisition & Pre‑processing (sensor streams, ERP data, document repositories)
   • LLM‑CA Core (model hosting, prompt engineering, inference orchestration)
   • Interaction & Augmentation (chat UI, voice interface, AR overlays)
   • Governance & Integration (security, versioning, API gateway, compliance checks)
3. Iterative Expert Review – The draft architecture was presented to three rounds of focus groups (≈10 participants each). Feedback was captured, prioritized, and fed back into the design.
4. Validation Criteria – Flexibility, scalability, modularity, and work‑augmentation potential were assessed qualitatively through the experts’ ratings and discussion notes.

Results & Findings

• High acceptance: Over 80 % of participants rated the architecture as “well‑aligned with daily QMS tasks.”
• Modularity confirmed: Experts highlighted that swapping out the LLM engine (e.g., from GPT‑4 to a domain‑specific model) would require changes only in the LLM‑CA Core layer.
• Scalability insights: The proposed API‑gateway and container‑orchestration strategy (Kubernetes‑style) was deemed sufficient for scaling from a single pilot line to multi‑plant deployments.
• Work augmentation: Simulated use‑cases (e.g., instant root‑cause analysis of a defect) showed a potential 30‑40 % reduction in time spent on manual documentation and knowledge retrieval.

Practical Implications

• Rapid prototyping for factories – Development teams can spin up a “LLM‑CA sandbox” using the architecture’s blueprints, connect it to existing MES/ERP data sources, and start testing assistant‑driven quality checks within weeks.
• Regulatory compliance made easier – Built‑in audit‑trail components automatically log prompts, model outputs, and user actions, helping manufacturers meet ISO 9001 or industry‑specific standards.
• Skill‑level agnostic interfaces – The interaction layer supports chat, voice, and AR overlays, allowing both seasoned quality engineers and line operators to benefit without steep learning curves.
• Future‑proofing – Because the architecture isolates the LLM engine, manufacturers can adopt newer, more efficient models (e.g., quantized or edge‑optimized LLMs) without re‑architecting the whole system.
• Open‑source potential – The component definitions and integration patterns could be shared as reference implementations, fostering a community around “AI‑augmented QMS” tools.

Limitations & Future Work

• Prototype‑level validation – The study stops at expert focus groups; real‑world pilot deployments are needed to quantify performance gains and uncover integration hiccups.
• Data privacy & IP concerns – While the architecture includes security hooks, the paper does not detail concrete encryption or on‑premises deployment strategies for highly confidential manufacturing data.
• Model bias & explainability – The authors acknowledge that LLM outputs may still contain hallucinations; future work will explore tighter grounding techniques and user‑friendly explanation modules.
• Scalability testing – Benchmarks under high‑throughput conditions (e.g., thousands of concurrent defect queries) remain to be performed.

Bottom line: This human‑centred architecture offers a concrete, developer‑friendly pathway to embed LLM‑powered assistants into quality management workflows, promising faster issue resolution, better knowledge capture, and a scalable foundation for the next generation of AI‑augmented manufacturing.*

Authors

• Marcos Galdino
• Johanna Grahl
• Tobias Hamann
• Anas Abdelrazeq
• Ingrid Isenhardt

Paper Information

• arXiv ID: 2603.16325v1
• Categories: cs.SE, cs.AI
• Published: March 17, 2026
• PDF: Download PDF