grubel
Grubel is a research lab automating the specialization loop that adapts AI models to individual legal matters. The platform jointly trains the model, agent, and harness on a matter's own documents at test time, enabling local deployment for data confidentiality. It targets high-stakes legal work where general-purpose agents fail on unstructured data, as demonstrated by its benchmark runs on Harvey's Legal Agent Benchmark.
- Legal Technology
- Artificial Intelligence
- AI Agents
- Software Only
Funding
Funding not disclosed
Founders
Founder details are not available yet.
Product
Problem
General-purpose AI systems are trained outside the context of individual legal matters, which each have their own world of facts, parties, customs, jurisdiction, and courts. When applied to critical legal work, these systems miss what the matter requires, and coding agents that are optimized for code and fixed environments fail silently on unstructured legal data.
Solution
Grubel automates the specialization loop that currently requires manual, repetitive labor — run, inspect, patch, run again — to adapt AI to a specific legal matter. Their platform brings data curation, adaptation, and evaluation into a continuous loop that trains and improves a model on the matter itself. The system specializes the model, agent, and harness jointly to the matter at test time, supported by a data engine and continual benchmarking. This approach aims to deliver frontier performance at lower cost, with matter-specific AI that can be deployed locally on the client's infrastructure in their jurisdiction.
Target Audience
Grubel targets legal professionals and firms handling high-stakes, complex matters that require jurisdiction-specific knowledge and data confidentiality, including those with large data rooms and extensive document review needs.
Features
- Automated specialization loop that jointly adapts model, agent, and harness to each legal matter at test time
- Data engine that performs curation and adaptation to train on the matter's own documents and facts
- Continual evaluation system that benchmarks improvements in a looping fashion
- Local deployment on client infrastructure, ensuring files never leave custody and ethical walls prevent data leakage between matters
- Research backed by practical evidence from benchmark testing, including runs on Harvey's Legal Agent Benchmark (LAB) showing performance gaps in general agents
- Proprietary approach to handling unstructured legal data, addressing failures of coding agents that rely on grep and keyword search methods