Ertas for Education

Fine-tune AI models on curriculum and anonymized course data, then run them on your institution's own infrastructure so student interactions stay with you.

The Challenge

Educational institutions at every level, from K-12 districts to research universities, are under pressure to personalize learning at scale. Adaptive tutoring, automated grading, curriculum recommendation, and content generation all benefit from AI, but the models need to understand the specific curriculum, pedagogical approach, and assessment standards of the institution. Generic language models produce plausible-sounding but often incorrect explanations, misalign with grade-level expectations, and lack awareness of specific course syllabi or institutional learning outcomes.

Student data privacy adds a formidable constraint. FERPA imposes strict requirements on how educational records are handled, and parents, students, and faculty are increasingly skeptical of sending student interaction data to third-party cloud AI providers. Many school districts have outright banned commercial AI tools that transmit student data externally. Institutions that want to use AI for grading assistance or tutoring must find a way to keep student data within their own systems while still accessing the benefits of modern language models.

The Solution

Ertas gives educational institutions a way to build curriculum-aware AI models and run them on their own infrastructure. In Ertas Studio, instructional designers and ed-tech teams fine-tune open models with LoRA on institution-specific content (course materials, past exam questions, rubrics and de-identified student interaction logs), using a visual canvas that runs training on Ertas's cloud GPUs. Faculty can review and approve training examples in Data Craft, Studio's guided dataset builder, so the model learns from explanations they stand behind. The resulting models use the institution's terminology, assessment criteria and teaching style.

Training data is uploaded to Ertas for training, so remove student names and identifiers before upload, and delete datasets from Studio when the project ends. Once trained, the model is exported from Hub as GGUF or safetensors and runs on the institution's own servers or private cloud, connected to the LMS through the institution's own integration. Student questions and the model's answers then stay inside systems the institution controls. Faculty and administrators get AI-powered tools (practice question generators, first-pass grading support and study recommendations) that run where the institution decides.

Key Features

Studio

Curriculum-Aware Fine-Tuning

Use Studio's visual canvas to fine-tune models on JSONL datasets of course content, assessment rubrics, Q&A pairs and teaching guidelines. Train subject-specific models (one for introductory biology, another for AP calculus) on small open models such as Phi-4 Mini or Llama 3.2 3B, and compare several runs side by side.

Studio

Faculty-Reviewed Examples

Data Craft, Studio's guided dataset builder, lets faculty write and edit tutoring exchanges as USER and ASSISTANT turns, version them, and track which examples are approved. The model trains on explanations your teaching staff have checked.

Hub

Models Your Institution Runs

Every trained model is listed in Hub for download as GGUF (llama.cpp, Ollama, LM Studio) or full safetensors (vLLM and other server runtimes). Host it on your own servers and connect it to Canvas, Moodle, Blackboard or a custom LMS through your own integration, with no Ertas dependency at runtime.

Hub

Student Interactions Stay Local

When the exported model runs inside your network, student questions and model responses stay in your systems, under your institution's own data policies. Training data is uploaded to Ertas's cloud GPUs, so anonymize student records before upload and delete datasets from Studio when you no longer need them.

Example Workflow

A large university's online learning team wants to provide adaptive tutoring for its introductory computer science course, which enrolls 2,000 students per semester. The team exports 15,000 anonymized student Q&A interactions from the course forum, along with excerpts from the course textbook and the assignment rubrics, as a JSONL dataset and uploads it to Ertas Studio. Teaching assistants use Data Craft to rewrite a few hundred answers in a Socratic style and approve them. In Studio, the team fine-tunes Phi-4 Mini with LoRA, training the model to guide students toward answers through questions instead of handing over solutions, and runs two configurations in parallel. They check both against a held-out set of forum questions graded by the teaching staff, then export the stronger model from Hub as GGUF and serve it with Ollama on a server in the university's data center. The course's Moodle instance connects to it through the university's own integration. Students get hints and explanations aligned with the course curriculum and grading rubric, and their conversations with the tutor stay on university infrastructure.

Compliance & Security

FERPA obligations, including what counts as an education record and who may access it, remain the institution's own work. Ertas changes where inference runs: after export, the model can run on institution-controlled servers, so student questions and model outputs stay there. Training works differently. Datasets are uploaded to Ertas's cloud GPUs for training and stored so runs can be repeated, and you can delete them. Anonymize or minimize student data before upload, and put the training set through your usual privacy review.