The Intent-based Servitization Manager translates high-level natural-language intents from MaaS stakeholders into structured, machine-executable JSON Requirement Models (RMs) aligned with the UniMaaS Formal Modelling Framework. Inputs are a textual user intent and a serialized provider/Knowledge-Graph constraint snapshot; outputs are three alternative, constraint-compliant RMs encoding goal, trigger, actions, constraints, resources, and optimisation profile. The component implements a three-stage pipeline : intent → RM generation; RM → manufacturing-process mapping via retrieve-and-match; deterministic selection and KG state update, built on a Mistral-7B-Instruct-v0.3 model, exposed through a full-stack prototype with a FastAPI backend, a React/TypeScript chatbot frontend, and a translator abstraction layer.
It supports all four UniMaaS pilots: AEGEAN, ADIENT, ANV, and CATONE with pilot-specific Pydantic validation schemas and is designed for fully local, self-contained deployment within the consortium’s Kubernetes-based testbed.

• Multi-candidate, constraint-bounded RM generation: the fine-tuned LLM produces three alternative JSON Requirement Models per intent, each conditioned on provider-specific numerical bounds, enabling downstream provider selection rather than committing to a single output
• Dynamic KG/provider-context conditioning at prompt level: the serialized provider-constraint snapshot is injected into the instruction prompt, exposing the model during fine-tuning to multiple state configurations for the same intent and enforcing operational feasibility at generation time.
• Translator abstraction layer: mock, prompt-based, and fine-tuned-model translators are interchangeable through a single configuration change, with no modifications to API routes, schemas, services, or frontend, enabling safe transition from interim to production translator and fully local LLM deployment.
• Unified pilot-agnostic architecture: one architecture covers four heterogeneous industrial domains (aircraft maintenance, automotive seats, additive manufacturing, warehouse logistics) via pilot-specific Pydantic schemas, intent guidelines, and KG terms, avoiding system duplication.
• Local, self-contained deployment of an open-source LLM: all inference runs on-premise on consortium infrastructure, eliminating data exposure to external APIs and addressing reviewer-raised security concerns.