Legal AI Platform: RAG over 10M+ Documents and Llama 3.1 Fine-Tuning

Summary

A UK law firm with 15+ lawyers wanted an AI system that works over its own legal corpus and is built on open-source models. I lead the project at SyntaxBrain, where I am Lead AI Engineer, as AI engineer and project manager: retrieval over more than 10 million documents, and fine-tuning of Meta’s open-source Llama 3.1 models on the firm’s domain.

Why RAG and fine-tuning together?

The project uses both techniques. Retrieval-augmented generation grounds every answer in the firm’s 10M+ document corpus, so answers can cite their sources. Fine-tuning Llama 3.1 adapts the model to the firm’s legal language and document formats. Retrieval supplies the facts; fine-tuning supplies the style and domain behaviour.

What I own

• Technical lead: architecture of the retrieval pipeline and the fine-tuning workflow for Llama 3.1.
• Project management: scope, milestones and communication with the client.

Why open-source models

The client chose open-source Llama 3.1 models as the base, so the models can be fine-tuned on the firm’s own legal domain.

Status

Client work under confidentiality. Architecture and results can be discussed in an interview.