Ertas for Legal

The Challenge

Law firms and corporate legal departments generate enormous volumes of contracts, briefs, memos, and case law research every day. Extracting patterns from this data (identifying unfavorable clauses in vendor agreements, summarizing depositions, or surfacing relevant precedent across thousands of filings) is exactly the kind of task large language models excel at. Yet generic models trained on internet text routinely misapply legal terminology, confuse jurisdictions, and produce citations to cases that do not exist.

The confidentiality problem is even more acute than the accuracy problem. Attorney-client privilege and work-product doctrine impose strict obligations on how legal data is handled. Sending client contracts or litigation strategy documents to a third-party API, where data may be logged, cached, or used for model improvement, creates privilege-waiver risks that no general counsel wants to accept. Many firms have blanket policies prohibiting the use of external AI services on client matters, leaving lawyers without access to the productivity gains the rest of the enterprise enjoys.

The Solution

Ertas lets legal teams build AI models that are accurate on their own domain. In Ertas Studio, legal technologists fine-tune open models on curated sets of anonymized contracts, case summaries, and internal knowledge management content, uploaded as JSONL or built in Data Craft, Studio's guided dataset builder, where lawyers review, edit and approve each example. LoRA fine-tuning keeps training efficient, and several candidate models can train in parallel for side-by-side comparison. The resulting models learn the firm's citation formats, contractual boilerplate, and jurisdiction-specific terminology.

Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas. Firms should anonymize or minimize it first and decide, under their own confidentiality obligations, which material is suitable for training. After training, the model is downloaded from Hub as GGUF or safetensors and runs on the firm's own servers, where the documents lawyers send it during day-to-day work stay inside the firm's network. Access, matter segregation, ethical walls and logging are handled by the firm's existing systems around that server. Firms that want help scoping a build can start with Ertas's enterprise custom-model service: a free use-case audit, then a three-to-four-week proof of concept.

Key Features

Example Workflow

A mid-size corporate law firm wants to speed up contract review for M&A due diligence. The legal technology team exports 30,000 commercial contracts from the firm's document management system, anonymizes them and removes client identifiers, and converts the clause-level annotations into a JSONL dataset. Senior associates spot-check a sample in Data Craft before it is uploaded to Ertas Studio. The team launches LoRA fine-tuning runs on Mistral 7B and Llama 3.1 8B in parallel, focused on clause classification and risk flagging, and compares each run against a held-out set of contracts the firm's lawyers have already reviewed. The better model is exported as safetensors and served with vLLM on the firm's own servers. Associates on the deal team send target-company contracts through an internal review tool connected to that server and get clause-by-clause risk notes on change-of-control provisions, indemnification caps and non-standard termination clauses. Lawyers review every flag, and their corrections feed the next training run.