Open Category Submission

621 – Bluebook AI Tool

The tool is a multi-agent software system designed to generate accurate legal citations from incomplete source inputs. It enables users to provide a list of sources, and specialized agents— dedicated to citation types such as case law, journal articles, and books—process each source independently. Each agent leverages APIs like Google Scholar and OpenLibrary to retrieve metadata, which is then used to construct Bluebook-compliant citations through an LLM-based agent.

Legal citation per The Bluebook's complex rules (about 400 pages of rules regarding legal citations) presents a significant challenge for law students, faculty, and practitioners. This challenge became evident through our interactions with law journal editors, law professors, and attorneys who reported spending dozens of hours finding missing information for citations and formatting them rather than engaging with actual legal work.
Our hypothesis is that a well-designed AI system can transform incomplete citation information into accurate Bluebook-compliant citations with minimal human intervention. Also, we understand that the users need a comprehensive solution that automates the complete Bluebook citation workflow. This is the reason why we decided to develop a tool that not only provides the Bluebook version of a citation, but is also able to read and parse Word documents to extract all citations, complete the missing information from the web, and reformat the citations in the correct way per the Bluebook rules. This end-to-end solution is a perfect example of a fully automated workflow.

Stream 3

Our system transforms incomplete footnote citations from .docx legal documents into accurate Bluebook-formatted citations using a modular, agent-based architecture powered by CrewAI. Each function in the workflow maps directly to a specific agent, with supporting technologies chosen for precision, scalability, and resilience.

The pipeline begins with footnote extraction via docx2python. The QueryExtractionAgent uses an LLM (Claude 3.5 via LangChain) to identify key search terms from raw footnotes. The ClassificationAgent determines whether a source is an Article, Book, or Case Law, using a combination of LLM reasoning and heuristics (is_case_citation()).

Sources are routed to the appropriate research agents — ArticleAgent, BooksAgent, or CaseLawAgent — which retrieve metadata using APIs (e.g., SerpAPI, OpenLibrary) and generate valid Bluebook citations through LLMs. These citations are validated using Pydantic models to ensure output consistency, with retry logic for malformed JSON. Finally, the FinalCitationAgent reintegrates the citation into the original footnote using context-aware LLM prompting.

This architecture is implemented in Python, using CrewAI for multi-agent orchestration, LangChain for LLM integration, Flask for backend orchestration, and AWS EC2 for deployment. Real-time updates are handled via a progress_emitter module, and MySQL is used optionally for persistent storage.

Each technology directly supports a functional layer — LLMs power semantic tasks, CrewAI manages agent workflows, and validation ensures legal-grade output integrity.

Our project offers a novel solution to a common and time-consuming problem in legal research: generating accurate Bluebook citations. Unlike existing citation tools that require users to manually input full metadata, our system can work from partial inputs (e.g., just a title or author) extracted from .docx footnotes — streamlining a typically manual, error-prone process.

The core innovation lies in the agent-based architecture powered by CrewAI, where each agent handles a specific part of the citation pipeline: query extraction, classification, metadata retrieval, citation formatting, and reintegration into footnotes. By decomposing the process into these modular agents, we ensure robustness, traceability, and the ability to retry or refine individual steps without rerunning the full pipeline.

We further enhance effectiveness by validating all AI-generated responses with Pydantic schemas and a retry mechanism. This allows us to maintain high output quality even when working with imperfect user input — a frequent challenge in real-world legal drafts.

Creatively, the tool doesn’t just “generate citations” — it reconstructs them in context. The FinalCitationAgent ensures that citations are embedded back into the original footnote in a natural and legally appropriate manner. This contextual awareness, guided by LLMs, is a major leap beyond rule-based formatters.
Together, these innovations allow us to turn incomplete, messy footnotes into professional-grade citations — automatically, scalably, and with minimal user effort. It’s a meaningful step forward for legal tech, especially for students, researchers, and professionals who frequently work with complex source material under time pressure.

Our solution is designed with scalability in mind, using a modular agent-based architecture that allows for parallel processing and efficient task distribution. Each step in the pipeline — from footnote parsing to citation generation — is handled by specialized agents running asynchronously. This enables horizontal scaling by distributing workloads across multiple instances or threads, particularly when deployed on cloud infrastructure such as AWS EC2.

To address potential bottlenecks, such as LLM response time or API rate limits, we implement request batching, output validation with retry control, and caching mechanisms for repeat queries. The architecture supports asynchronous processing, allowing the system to handle multiple documents and users concurrently without significant performance degradation.

To foster long-term engagement, we plan to build features like citation history, customizable formatting templates, and integrations with citation managers (e.g., Zotero) or legal writing platforms. Our modular design allows for continuous upgrades, such as supporting more citation formats or integrating OCR for scanned documents.

We will gather ongoing user feedback to adapt to changing academic and legal standards, ensuring the tool evolves alongside user expectations. By combining scalable infrastructure, efficient design, and user-centered development, we aim to deliver a sustainable and impactful legal research tool that meets growing demand while remaining accessible and environmentally conscious.

Our Bluebook citation generator addresses efficiency and educational challenges in the legal profession, redirecting valuable human attention from technical formatting to substantive legal work.
For law journal editors who volunteer their time, Bluebook citation review represents an enormous burden. Student editors often spend 60-70% of their journal time verifying and correcting citations rather than engaging with legal scholarship. Our tool dramatically reduces this technical overhead, allowing editors to focus on evaluating legal arguments, improving article structure, and providing meaningful feedback to authors. This shift from technical formatting to content analysis improves both the educational experience for editors and the quality of legal scholarship.
For small law firms and solo practitioners without dedicated support staff, our solution democratizes access to professional-grade citation capabilities. While large firms can delegate citation tasks to specialized staff, smaller practices must handle these details themselves, creating an efficiency gap. Our tool levels this playing field, allowing practitioners at firms of all sizes to dedicate more time to client service and legal strategy rather than technical formatting work.
For legal interns and students, whose learning opportunities are precious and limited, our tool transforms how they spend their time. Rather than spending hours formatting citations—a task with limited educational value—they can focus on substantive legal analysis, client interaction, and other high-value professional development activities. This ensures that educational time is spent on skills that truly advance their legal capabilities.
To measure social impact, we will track:
Time saved on citation tasks
Adoption rates among the legal community, especially among the above mentioned groups.
We will maintain responsiveness through regular feedback sessions with journal editorial boards, small firm practitioners, and legal education directors to ensure our tool evolves with their needs.

Personal Information Collection Statement (PICS):
1. The personal data collected in this form will be used for activity-organizing, record keeping and reporting only. The collected personal data will be purged within 6 years after the event.
2. Please note that it is obligatory to provide the personal data required.
3. Your personal data collected will be kept by the LTTC and will not be transferred to outside parties.
4. You have the right to request access to and correction of information held by us about you. If you wish to access or correct your personal data, please contact our staff at lttc@eduhk.hk.
5. The University’s Privacy Policy Statement can be access at https://www.eduhk.hk/en/privacy-policy.

• I have read and agree to the competition rules and privacy policy.