Researcher Anisha Jadhav’s New LLM- Based Framework Aims to Bring Transparency to Agile Story Point Estimation

Anisha Jadhav's SPX framework introduces explainable AI to Agile story point estimation, moving beyond black-box predictions. This IEEE-published research helps software teams understand 'why' estimates are made, fostering transparency and trust. It's a game-changer for confident project planning, ensuring AI supports better decision-making, not just automation.

California-based IEEE-published researcher Anisha Jadhav’s SPX framework signals a new direction for AI in software engineering: transforming Agile story point estimation from a black-box prediction into an explainable planning process.

As artificial intelligence continues to reshape software development, much of the attention has focused on tools that write, review, or generate code. But AI/ML researcher Anisha Jadhav is examining a different challenge, how software teams plan the work before coding begins.

Jadhav’s SPX work was published as part of the IEEE conference record in 2025. The paper, "SPX: A Novel LLM-Based Framework for Explaining and Estimating User Story Points," appeared through the IEEE 3rd International Conference on Artificial Intelligence, Blockchain, and Internet of Things, known as AIBThings 2025, in Mount Pleasant, Michigan. Read the paper on IEEE Xplore here. The publication places Jadhav’s research within a growing conversation about how AI should be used in software teams, not just to generate code, but to support the planning decisions that shape delivery timelines.

Jadhav’s work spans machine learning, deep learning, software engineering, data analytics, and LLM-driven automation. Her research includes multiple IEEE-indexed publications across artificial intelligence, intelligent systems, software engineering, computer vision, and applied machine learning, alongside industry experience building data systems and decision-support dashboards.

In Agile development, story points are used to measure the relative effort, complexity, and uncertainty involved in completing a user story. These estimates influence sprint planning, resource allocation, delivery timelines, and team velocity. Yet the process often depends heavily on human judgment. Traditional methods such as planning poker and team consensus can encourage discussion, but they can also vary widely across teams, projects, and technical environments.

Jadhav’s SPX framework aims to make that process more transparent. Rather than producing only a numerical estimate, SPX generates a plain-language explanation describing why a user story may require a certain level of effort. The explanation may point to factors such as technical complexity, unclear requirements, dependencies on external systems, integration challenges, or implementation risk.

That focus on explainability is what separates SPX from many previous approaches to software effort estimation. Earlier machine learning and deep learning models have shown promise in predicting story points, but many operate as black boxes. They may estimate whether a task should receive three, five, or eight points, but they often do not explain the reasoning behind the prediction. In real Agile environments, that lack of transparency can make AI-assisted estimation difficult to trust.

"Across domains, whether in software engineering or enterprise planning, AI systems must be verifiable," Jadhav said. "A blind prediction cannot drive high-stakes technical decisions."

SPX combines several artificial intelligence techniques, including GPT-4-based few-shot learning, Sentence-BERT embeddings, active learning, contextual similarity matching, and natural language explanation generation. The system analyzes the text of a user story, compares it with relevant historical examples, and generates both an estimated story point value and a human-readable rationale.

According to the research, SPX was evaluated on more than 6,000 user stories collected from open-source Jira repositories, including projects such as Apache, Spring, and JBoss. The framework achieved a mean absolute error of 1.42, outperforming several baseline methods, including traditional machine learning models, deep learning approaches, transformer-based estimators, and ensemble systems.

But the central contribution of Jadhav’s work is not only accuracy. It is transparency.

In a human evaluation involving Agile practitioners, SPX-generated explanations received an average clarity score of 4.3 out of 5, with strong ratings for relevance and justification. That result suggests the system’s explanations were not only generated by an AI model but also understandable to people familiar with Agile planning.

Software engineering experts say that explainability is becoming increasingly important as AI enters planning and management workflows. In agile environments, they note, teams are more likely to adopt AI-assisted tools when those systems can show not only a prediction but also the reasoning behind it.

For software teams, that distinction matters. Sprint planning is rarely a purely numerical exercise. Developers, product owners, scrum masters, and engineering managers often need to discuss why one task is straightforward while another carries hidden complexity. If an AI system recommends a higher estimate, teams need to know whether the reason is vague acceptance criteria, technical debt, dependency risk, or unfamiliar implementation details.

Jadhav’s framework also includes an active learning component that flags uncertain or ambiguous user stories for human review. This keeps human expertise involved where it is most needed, while allowing the system to improve over time. Another feature, contextual analogy matching, compares new stories with similar past examples, reflecting how experienced Agile teams often estimate work based on previous tasks.

As first author, Jadhav positions SPX within a growing area of research exploring how large language models can support software engineering beyond code generation. The focus is decision support: helping teams plan, estimate, and reason about work more clearly.

The research also acknowledges limitations. SPX was evaluated on public datasets, which may not fully reflect proprietary enterprise environments or company-specific estimation practices. Future work would need to examine enterprise deployment, team-specific calibration, real-time performance, and integration with tools such as Jira.

Still, Jadhav’s research raises a timely question for the software industry: can AI help teams not only build faster but also plan better?

SPX suggests that the future of AI in software engineering may extend beyond coding assistants. It may also include explainable planning tools that help teams understand complexity, identify uncertainty, and make more confident development decisions.