Ertas for Reusable Client Data Pipelines
For AI/ML service providers: keep versioned, reusable training datasets in Data Craft, repeat the same training setup in Ertas Studio for each engagement, and hand every client a separately trained model they can own and run on their own servers.
The Challenge
AI/ML service providers rebuild data preparation pipelines for every client engagement. Different scripts, different tools, different quality standards. No standardization means no institutional knowledge captured. Each engagement's pipeline is a one-off that dies when the project ends.
The Solution
Ertas gives service providers one repeatable path from client examples to a trained model the client owns. Data Craft, Studio's guided dataset builder, holds each engagement's training data as conversational examples: the team adds USER and ASSISTANT turns, edits and versions them, and tracks which examples the client's domain expert has approved. A dataset structure that worked for one client, such as the turn format and the kinds of examples it covers, becomes the starting point for the next engagement, so the team fills it with the new client's material instead of starting from scratch.
Training follows the same steps each time. The team picks an open base model in Studio, configures LoRA fine-tuning on the visual canvas and runs it on Ertas's cloud GPUs, with several runs in parallel to compare settings or model sizes. Datasets are stored, so a run can be repeated later on the same data. Each client gets its own model, exported from Hub as GGUF or full safetensors, which the client can own and run on its own servers with no Ertas dependency. For clients who want help scoping the work, Ertas also runs enterprise custom-model builds that start with a free use-case audit.
Key Features
Versioned Datasets in Data Craft
Build each client's training set as USER and ASSISTANT turns in Data Craft. Edit and version examples as the engagement goes, and keep the structure that worked as the template for the next client.
Expert Review and Approval
Track approval status on every example, so the client's domain experts sign off on what the model learns before a training run starts. The same review step carries across every engagement.
Repeatable Training Runs
Configure LoRA fine-tuning of an open model on Studio's visual canvas and run it on Ertas's cloud GPUs. Run several configurations in parallel, compare them, and rerun the same setup when the client sends new data.
A Model Each Client Owns
Every engagement ends with its own trained model in Hub. Export it as GGUF for llama.cpp, Ollama or LM Studio, or as full safetensors for vLLM, and hand it over. The client runs it on its own servers with no Ertas dependency at runtime.
Example Workflow
A service provider builds a support-assistant dataset in Data Craft for a healthcare client: a few hundred USER and ASSISTANT examples covering appointment questions, referral steps and clinic policies, written from de-identified source material and approved by the clinic's practice manager. They train two LoRA runs in parallel in Ertas Studio on Qwen 2.5 3B and Llama 3.1 8B, compare both against a held-out set of real questions, and export the better model from Hub as GGUF for the clinic to run on its own server. The next engagement is a legal client. The team reuses the same dataset structure and review process, fills it with the firm's matter-intake questions, has a senior associate approve the examples, and repeats the same training setup. Two clients end up with two separately trained models, each owned by the client that paid for it, built on one workflow the provider improves with every engagement.
Related Resources
Ship AI that runs on your users' devices.
Start free with 150 credits after card verification. Paid plans from $20/mo USD.