Fetch legal issue classifer
FETCH is a lightweight, ensemble legal-issue classifier that helps intake systems classify a user's legal issue and route them to the right service, faster and at lower cost than a single large model.
Project Description
The FETCH (Fast Ensemble Tagging and Classification Helper) classifier is an AI-powered tagging system built to improve intake and referral workflows in legal services. Instead of relying on one model, FETCH uses a weighted, multi-label ensemble: several small LLMs (e.g., GPT-5-nano, gemini-2.5-flash, mistral-small), a traditional ML model, and keyword rules vote on the best labels from a 200+ node taxonomy. This design delivers practical speed and cost savings while maintaining high labeling accuracy for real-world intake text (phone notes, web forms, emails).
FETCH is designed for plug-in use by legal organizations’ phone trees and web intakes. Deployments with partners such as Virginia Legal Aid Society and Oregon State Bar show how incoming messages can be auto-classified to issue codes and routed to the appropriate unit, panel, or referral partner. When the ensemble’s confidence is low, FETCH auto-generates targeted follow-up questions to clarify jurisdiction, parties, or problem type before making a routing decision.
The project was designed and led by Quinten Steenhuis with collaborators at Suffolk University Law School, Lemma Legal, and partner organizations. A 2025 write-up credits contributors including David Colarusso, Jack Adamson, and Jim Graszer, and reports evaluations that compare ensemble outputs to single-model baselines on legal-aid intake data, highlighting gains in throughput, cost control, and stable accuracy across a large label set.
Read more about FETCH's development and evaluation here at this paper.