Automate Law Firms "Fund Review" processes using AI Document intelligence and Generative AI

Abstract

The increasing volume and complexity of legal and financial documents require a shift from manual review to intelligent automation. However, current AI paradigms present critical risks in this field. Traditional supervised models fail due to the scarcity of large-scale, paragraph-level labeled data, while domain-specialized transformers like FinBERT or LegalBERT depend on fine-tuning over datasets that are unavailable in our domain. Indeed, as they are shaped by their training corpora, they do not reflect the specific structure, semantics, and regulatory nu- ances of fund prospectuses, constitutions, and agreements. In fact, there are no large-scale, legal paragraph-level datasets for our corpus that map legal texts into predefined legal categories. Without this supervision, classifiers can’t be trained reliably. This also implies that there is no taxonomy available in our domain, which is crucial for generating fund review reports. Large Language Models (LLMs) such as GPT-5, despite their impressive general reasoning abilities, have broad and ungrounded knowledge. In a legal and financial context, relying solely on an LLM’s internal representation introduces risk: it cannot explain its reasoning paths or verify its sources. Therefore, this thesis introduces a different philosophy: an hybrid architecture for the auto- mated analysis, synthesis, and querying of legal corpora by leveraging the best of both worlds: a knowledge graph (KG) that acts as the system’s "legal brain", and an LLM leveraged not as a source of truth, but as a sophisticated language interface to interact with the KG. Instead of relying on an LLM’s internal knowledge, we externalize all legal reasoning into this symbolic structure. By grounding LLM outputs in this graph, we ensure that retrieval, summarization, and drafting of legal texts are legally valid, context-aware, and aligned with the structural logic of the legal corpus. The framework’s efficacy is demonstrated through three key achievements: (1) the con- struction of a knowledge graph from a corpus of legal documents, using structure-aware parsing algorithms that preserve legal semantics; (2) the automated discovery of an emergent legal taxonomy using graph-based community detection on semantic similarity links, eliminating the need for predefined classification schemes that aims to write fund review reports; and (3) the implementation of a Graph-RAG (Retrieval-Augmented Generation) system that queries the KG to provide accurate, domain-aware, and legally interpretable texts.