Custom AI Models for Legal Document Review, Run on Your Firm's Servers
Ertas lets law firms and legal departments train custom models for contract analysis and document review on their own work product, then run them on the firm's own servers, where the documents they review in production stay in-house.
The Challenges You Face
Client Confidentiality Shapes Every AI Decision
Uploading client documents to third-party AI services raises questions under professional confidentiality duties and client engagement terms. Many firms restrict which tools may touch client material, which rules out sending every document to a hosted API.
Legal Documents Require Domain-Specific Understanding
Generic LLMs misread legal terminology, miss jurisdictional nuances and produce unreliable contract analysis. Legal language has precise meanings that differ from common usage, and a model trained on general text cannot reliably tell them apart.
Document Review Is the Largest Cost Center
Large-scale document review for litigation, due diligence and regulatory work consumes thousands of attorney hours. AI-assisted review could cut that cost sharply, but only if the AI understands the specific legal context well enough to be trusted.
AI-Assisted Work Needs to Be Explainable
When AI assists in legal work, firms need to show what the model was trained on, which examples were reviewed, and that a lawyer checked the output. Ad-hoc tools make that hard to document.
How Ertas Solves This
Ertas Studio trains a custom model on your firm's own work product: contracts with the clauses your associates flag, memos in your house structure, review decisions from past matters. Lawyers and paralegals can build, correct and approve examples in Data Craft, Studio's guided dataset builder, with no code. Training runs on open models through a visual canvas, and several runs can go in parallel so you compare them before choosing.
Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas. Choose training material your engagement terms allow you to use this way, redact names and identifiers first, and delete datasets when you no longer need them.
The trained model belongs to the firm. Download it as GGUF or full safetensors and run it on the firm's own servers with llama.cpp, Ollama or vLLM. Documents reviewed in production stay inside the firm's network, with no Ertas dependency at runtime. Your professional obligations remain a matter for your own judgment, and running review models in-house keeps client documents within systems you already control.
Key Features for Law Firms & Legal Departments
Lawyer-Reviewed Training Data
Data Craft presents each example as an exchange your lawyers can edit, version and mark approved, so the training set carries the same judgment they apply in manual review.
Practice-Specific Model Training
Fine-tune open models on your firm's own work product, such as contract analysis patterns, memo structures and review decisions, so the model reflects your firm's standards.
Run on the Firm's Servers
Trained models are listed in Hub and download as GGUF for llama.cpp or Ollama, or as full safetensors for vLLM. Serve them on hardware the firm controls, where client documents stay in-house.
Versioned Training Sets
Data Craft versions your examples and tracks approval status, so you can show which approved examples went into each model version.
Why It Works
- On the QMSum meeting-summarization benchmark, a 1.2B model trained on Ertas scored 35.41 ROUGE-1 against 32.43 for the best of five hosted frontier models: a small model trained for one task can beat far larger general models on that task.
- Exported models run on llama.cpp, Ollama, LM Studio or vLLM on hardware the firm controls, so production inference needs no connection to Ertas.
- Lawyers contribute directly in Data Craft, a conversational editor for building, versioning and approving training examples.
Example Workflow
A corporate law firm reviews thousands of contracts in due diligence engagements and wants a model that flags change-of-control, indemnification and non-compete provisions. A knowledge management lawyer assembles 2,000 clause examples from closed matters the firm's engagement terms allow it to reuse, with client names redacted, and formats them as a JSONL dataset.
Senior associates review the examples in Data Craft, correct the clause labels and approve the set. The litigation support team trains runs on two base models in parallel in Ertas Studio and compares them against a held-out set of clauses the associates labeled separately.
The better model is downloaded as a GGUF file and served on the firm's own GPU server. On the next engagement it marks likely provisions for junior associates, who verify every flag before anything reaches the client.
Related Resources
Ship AI that runs on your users' devices.
Free plan with 30 credits/mo, no card required. Paid plans from $10/mo USD.