Custom AI Models for Clinical Text, Run on Your Own Servers
Ertas Studio lets healthcare teams train custom models on their own clinical examples and export them to run on hospital-owned servers, so the patient data a model reads in production stays inside your network.
The Challenges You Face
Patient Data Raises the Bar for Cloud AI
HIPAA, state privacy laws and institutional review boards put strict conditions on where protected health information goes. Sending every clinical note to a third-party AI API means vendor agreements, risk reviews and ongoing exposure for each new use case, which stalls many projects before they start.
Clinical Language Trips Up General Models
Clinical notes are full of abbreviations, shorthand and local conventions that general-purpose models misread. A model that has never seen how your clinicians write produces extractions and summaries that someone has to check line by line.
Per-Token API Costs Grow With Patient Volume
A hospital produces thousands of notes, reports and letters a day. Paying a frontier model per token for every one of them turns a promising pilot into a budget line that grows with patient volume.
Domain Expertise Lives with Clinicians
The people who understand clinical workflows, medical terminology and patient context are clinicians, and few of them write code. Healthcare AI improves fastest when clinicians can review and correct training examples directly.
How Ertas Solves This
Ertas Studio trains a custom model on examples of the task you want done: extracting medication lists, coding encounters, drafting discharge summaries. Clinicians review, correct and approve those examples in Data Craft, Studio's guided dataset builder, without touching code. You choose an open model such as Llama, Qwen, Gemma or Phi, train it through a visual canvas, and run several training runs in parallel to compare them.
Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas. Build it from records your organization has already de-identified or minimized. Datasets are stored so runs can be repeated, and you can delete them at any time.
Once trained, the model is yours. Export it as GGUF for llama.cpp, Ollama or LM Studio, or as full safetensors for vLLM, and run it on your hospital's own servers with no Ertas dependency at runtime. Every note the model reads in production stays inside your network. Meeting HIPAA and your institution's policies remains your organization's own work, and running inference in-house keeps production patient data out of another vendor's systems.
For larger programs, Ertas also runs enterprise custom-model builds: a free use-case audit, a three-to-four-week proof of concept, the build, then running and retraining, with your organization owning the trained model.
Key Features for Healthcare Organizations
Clinician-Reviewed Training Examples
Data Craft lets clinicians add and edit example exchanges as USER and ASSISTANT turns, version them, and mark each one approved. The training set reflects clinical judgment, and nobody has to learn developer tools to contribute.
Visual Training on Open Models
Pick an open base model, attach your dataset and launch LoRA fine-tuning on Ertas's cloud GPUs from a visual canvas. Run several configurations in parallel and compare the results before you choose one.
Export and Run In-House
Your trained models are listed in Hub. Download each one as GGUF for llama.cpp, Ollama or LM Studio, or as full safetensors for vLLM, and serve it on hospital-owned hardware so inference stays inside your network.
Adaptive Model Endpoint
For a single high-volume task, the adaptive model endpoint gives you one API. A custom model trained on your own requests and your team's feedback takes over a growing share once it passes your evals, while hard cases stay on the frontier model you already use. It runs Ertas-managed, in your own cloud, or on-premise.
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 you control, so production inference needs no connection to Ertas.
- Clinicians contribute directly in Data Craft, a conversational editor for building, versioning and approving training examples, with no code required.
Example Workflow
A hospital's informatics team wants a model that extracts medication names, doses and frequencies from unstructured clinical notes. A data engineer takes 2,000 notes that the hospital has already de-identified under its own process and loads them into Data Craft as example exchanges: the note as the USER turn, the structured medication list as the ASSISTANT turn.
A small group of clinicians reviews the examples in Data Craft, corrects the extractions and approves each one. The team then trains two runs in parallel in Ertas Studio, one on Qwen 2.5 7B and one on Llama 3.1 8B, and compares both against a held-out set of notes the clinicians labeled separately.
The better run is exported as a GGUF file and served with llama.cpp on the hospital's own GPU server. New notes are processed inside the hospital network, and clinicians spot-check a sample of outputs each week, adding corrected cases to Data Craft for the next training run.
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.