Ertas for Healthcare

The Challenge

Healthcare organizations sit on vast amounts of unstructured clinical data (physician notes, discharge summaries, radiology reports, and patient communications), yet extracting actionable intelligence from this data remains painfully difficult. Generic large language models hallucinate medical terminology, misinterpret clinical abbreviations, and lack the nuanced understanding of disease pathways that clinicians rely on every day.

The regulatory landscape makes the problem even harder. HIPAA, HITECH, and emerging state-level privacy laws impose strict requirements on how protected health information (PHI) is stored, processed, and transmitted. Sending patient data to third-party API endpoints is a non-starter for most compliance teams, and the lack of on-premise or private-cloud inference options from major AI providers forces hospitals and health-tech companies into an impossible tradeoff between innovation and compliance.

The Solution

Ertas gives healthcare teams a way to build domain-specific models they own and can run on their own infrastructure. In Ertas Studio, clinical NLP engineers upload a de-identified JSONL dataset, or build one in Data Craft, and fine-tune open models such as Llama, Qwen, Gemma or Mistral with LoRA. The resulting models learn your ICD codes, SNOMED CT concepts and clinical shorthand from your own examples. Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas: de-identify it to your organization's standard before upload, and delete datasets from Studio when you no longer need them.

Once a model passes your evaluation, you export it from Hub as GGUF or safetensors and run it on servers you control, with Ollama, llama.cpp or vLLM inside your VPC or on-premise cluster. Inference requests and outputs stay on those servers, and the model has no dependency on Ertas at runtime. Access controls, logging and the rest of your HIPAA program stay with your own security team, using the tools they already run. For organizations that want help scoping a clinical use case, Ertas's enterprise team offers a free use-case audit and a three-to-four-week proof of concept, and the organization owns the trained model.

Key Features

Example Workflow

A hospital's clinical informatics team de-identifies 50,000 radiology reports under its own de-identification procedure and exports them as a JSONL dataset of findings paired with impressions. A radiologist reviews a sample in Data Craft, corrects a few impressions and approves the examples the team will train on. The team uploads the dataset to Ertas Studio and runs LoRA fine-tuning on two base models in parallel, Mistral 7B and Qwen 2.5 7B, targeting radiology impression generation. They compare both runs against a held-out set of reports scored by their own radiologists and pick the stronger model. They export it from Hub as safetensors, serve it with vLLM on a server inside the hospital network, and expose it only to the EHR integration layer. Radiologists see draft impressions that they edit and sign, and every inference request runs on hospital hardware. When the review is done, the team deletes the training dataset from Studio.