Use Case Points-Based Software Effort Estimation Using Multimodal Llms For UML Use Case Diagram Analysis With Langgraph

Abstract

Estimating software development effort during early project stages using Karner's Use Case Points (UCP) method has traditionally depended on subjective human judgment to interpret visual UML diagrams and calibrate environmental parameters. The recent progress in Multimodal Large Language Models (MLLMs) provides powerful visual reasoning abilities, yet there are serious reliability risks in predicting project effort directly from an image. Standard UML use case diagrams illustrate visual boundaries and functional relationships, but they do not contain organizational metadata such as developer experience, team motivation, or hardware environment stability. When vision models are asked to guess these unseen parameters directly from visual artifacts, they produce unreliable hallucinations. To address this methodological gap, this paper proposes a hybrid, agentic framework based on LangGraph, which explicitly decouples the extraction of visual diagrams from deterministic mathematical calculation. Our pipeline strictly partitions responsibilities: an MLLM agent is restricted exclusively to parsing the visual elements of UML diagrams, identifying actors and extracting explicit use case transaction counts to calculate the Unadjusted Actor Weight (UAW) and Unadjusted Use Case Weight (UUCW). Concurrently, Technical Complexity Factors (TCF) and Environmental Complexity Factors (ECF) are supplied as validated user constraints through structured Pydantic schemas, preventing model hallucination of non-visual data. We evaluated our system across a benchmark of 25 annotated UML use case diagrams representing diverse software domains. The framework executed deterministically across 24 completed diagrams, achieving 100% extraction precision and recall against ground truth. When tested on a diagram deliberately missing transaction details, the system activated a built-in safety guardrail, halting execution instead of fabricating numbers. By separating visual perception from contextual project variables, this framework delivers an auditable, practical pathway for early-stage software cost estimation.