Case StudyAIWeb

snowLEX: AI Legal Research Platform

A legally reliable AI research platform combining large-scale vector search with a purpose-built legal dashboard for real-world legal workflows.

Visit Project→

About

snowLEX was built to solve two deeply connected problems in modern legal technology: the unreliability of generic AI in legal reasoning, and the lack of usable, modern interfaces for AI-powered legal research. The platform enables lawyers and researchers to interact with AI through a structured dashboard: managing chats, uploading legal documents, reviewing citations, and navigating sources with full transparency. Every UI decision was guided by legal usability principles — clarity, traceability, and trust. snowLEX treats law as structured knowledge, not just text. The system enforces citation grounding at every step, ensuring that AI responses are not only fast, but legally defensible.

The Challenge

The core challenge was engineering both intelligence and usability at scale. On the backend, over 1.5 million EU legal documents were statistically overpowering Finnish national statutes, causing incorrect prioritization in search results. On the frontend, most AI legal tools fail to offer interfaces suitable for professional legal work. The challenge was to design a system that retrieves the right law (not just the closest text), guarantees traceable citations, and provides a dashboard lawyers can actually work with daily.

Features

Hybrid search engine indexing 2M+ legal vectors

RAG pipeline with strict source verification

Zero-hallucination citation enforcement

Full-featured Next.js legal dashboard

Multi-chat system with session persistence

Document upload & analysis (PDFs)

Metadata-aware filtering (jurisdiction, year, authority)

Local LLM inference via Ollama for privacy

Journey

Phase 1

Legal Research & Feasibility Study

Extensive research on what authoritative legal data is available from Finlex and EUR-Lex, understanding document structures, metadata quality, and why generic AI systems fail in legal contexts.

Phase 2

MVP Architecture & Early Prototyping

A basic RAG pipeline was implemented to test whether retrieval-augmented generation could work reliably for legal queries. Retrieval logic was prioritized over generation quality.

Phase 3

Embedding Pipeline & Vectorization

Built a large-scale embedding pipeline vectorizing legal documents from Finlex and EUR-Lex with rich metadata structures (year, authority, jurisdiction, document type).

Phase 4

Experimental Interface (Streamlit)

Streamlit was used as a research tool to rapidly test legal queries, visualize retrieved sources, and validate citation behavior before committing to a full web product.

Phase 5

Transition to Full Web Platform

Streamlit was replaced with a full Next.js web application featuring a professional legal dashboard, chat-based research workflows, document upload, and multi-session support.

Phase 6

Large-Scale Data Expansion & System Hardening

The dataset expanded to 2M+ legal documents. Retrieval logic was refined to prevent EU data from drowning out Finnish law. Zero-hallucination constraints were enforced more strictly.

Impact

90% reduction in legal research time

2M+ legal documents embedded and indexed

100% citation-backed answers (zero hallucination tolerance)

Full-featured legal research dashboard adopted for real workflows

Tech Stack

Next.jsTypeScriptPythonFastAPIQdrantOllamaFirebaseGoogle CloudBGE-M3 Embeddings

Visit Project→View All Projects