Legal professionals routinely spend significant time on case analysis tasks that follow well-defined reasoning protocols, of which the Issue-Rule-Application-Conclusion (IRAC) method is the most widely taught and practiced framework across common-law jurisdictions. Within IRAC, the calculation of damages is particularly demanding: it requires the lawyer to identify the correct cause of action, retrieve the governing statutory and case-law rules (such as the Hadley v Baxendale test for remoteness in contract, the once-and-for-all rule for tort damages, and jurisdiction-specific statutory caps and multipliers), gather and quantify the relevant facts (lost profit, mitigation, contributory conduct), apply mathematical formulae correctly, and justify each step against authority. Large language models (LLMs) have shown promise on surface-level legal reasoning but fail in two critical ways for this setting. First, off-the-shelf LLMs reach incorrect conclusions on roughly half of IRAC scenarios and rarely align with the structured reasoning experienced lawyers produce [1]. Second, while integrating LLMs with a semi-structured legal knowledge base substantially improves issue identification and bridges the gap between legalese and everyday language [2], current systems remain weak at the quantitative Application step that underpins damages assessment, which combines symbolic rule application with arithmetic reasoning, multi-step tool use, and verification against precedent. 

This project aims to design and build an AI agent that automates end-to-end IRAC analysis for legal professionals, with damages calculation as its central capability. The student will:

• Develop a multi-agent architecture that decomposes IRAC into specialised agents for issue spotting, rule retrieval, fact extraction, application and calculation, and conclusion drafting, with a critic agent that audits each step against the retrieved authorities.
• Build a calculation-grounded reasoning module that combines LLMs with symbolic tools (formula libraries, numerical solvers, units and currency handling, citation checkers) so that every damages figure is traceable to facts, rules, and explicit arithmetic operations rather than fabricated through free-form generation.
• Extend the semi-structured legal knowledge base from [2] to cover damages-specific authorities, including statutory caps, indexation tables, interest rates, leading precedents on heads of loss, and the conditional and cross-referenced structure characteristic of damages law, and develop retrieval methods that can navigate this structure reliably.

The goal is an AI legal-reasoning agent that substantially reduces the time legal professionals spend on routine damages analysis while keeping every conclusion verifiable, well-cited, and aligned with the IRAC protocol they already trust.

[1] Xiaoxi Kang, Lizhen Qu, Lay-Ki Soon, Adnan Trakic, Terry Yue Zhuo, Patrick Charles Emerton, and Genevieve Grant. Can ChatGPT Perform Reasoning Using the IRAC Method in Analyzing Legal Scenarios Like a Lawyer? In Findings of EMNLP, pages 13900–13923, 2023.

[2] Xiaoxi Kang, Lizhen Qu, Lay-Ki Soon, Zhuang Li, and Adnan Trakic. Automating IRAC Analysis in Malaysian Contract Law Using a Semi-Structured Knowledge Base. Artificial Intelligence and Law, pages 1–44, 2025.