Ertas for Customer Support

Fine-tune AI models on your real support conversations so they resolve tickets accurately, in your product's own terms.

The Challenge

Customer support teams are drowning in ticket volume while struggling to maintain quality. Generic AI chatbots promise relief but consistently disappoint: they hallucinate product features that don't exist, give confidently wrong troubleshooting steps, and escalate to human agents at rates so high that the automation barely moves the needle. The root cause is simple: these models have never seen your product documentation, your internal runbooks, or the thousands of nuanced edge cases your senior agents have learned to handle.

The problem deepens as products grow more complex. Every new feature, pricing change, or integration adds another surface area where a generic model will fail. RAG pipelines help but introduce their own reliability issues: chunking artifacts, retrieval misses, and context window overflows that degrade answer quality unpredictably. Support leaders need AI that genuinely understands their product at the depth of a trained agent, not a system that pattern-matches against retrieved snippets and hopes for the best.

The Solution

Ertas turns your best support knowledge into a purpose-built AI model. In Studio, support operations teams fine-tune open models on resolved ticket histories, internal knowledge bases and curated Q&A pairs, uploaded as JSONL or built in Data Craft, Studio's guided dataset builder, where senior agents edit and approve example conversations. The resulting model has learned your product's logic, terminology and troubleshooting paths, so its answers read like your best agents' answers.

Each trained model is listed in Hub for export as GGUF or full safetensors. Serve it with Ollama, llama.cpp or vLLM on your own infrastructure and connect it to your helpdesk platform through your own API, with no Ertas dependency at runtime. Models in the 3B to 8B range respond quickly on modest GPUs. As your product changes, retrain in Studio on fresh ticket data, compare the new run against the current one, and swap in the exported model once it holds your quality bar. Teams that would like this run as a service can use the adaptive model endpoint: a custom model trained on your own requests and your team's feedback takes over a growing share of support requests once it passes your evals, while hard cases stay on the frontier model you already use.

Key Features

Studio

Support-Tuned Model Training

Import resolved tickets, macros and knowledge base articles as JSONL training data into Studio. Fine-tune with LoRA so the model learns your product's terminology, common failure modes and preferred resolution paths, and run several configurations in parallel to compare them.

Studio

Conversational Base Models

Start from instruction-tuned open models in Studio's catalogue, including Llama 3.1 8B, Qwen 2.5 7B, Mistral 7B and Gemma 3. Pick a smaller model such as Llama 3.2 3B for simple FAQ answers and an 8B model for complex troubleshooting.

Hub

Low-Latency Serving

Export the trained model from Hub as a quantized GGUF for llama.cpp or Ollama, or as safetensors for vLLM, and serve it on your own hardware close to your helpdesk. You choose the hardware, the scaling and the rollout process, including a staged rollout when you swap in a retrained model.

Hub

Conversation Data Protection

When you serve the exported model yourself, live customer conversations stay inside your network. Training data is uploaded to Ertas's cloud GPUs, so remove email addresses, phone numbers and account IDs from tickets before upload, and delete datasets from Studio when they are no longer needed.

Example Workflow

A SaaS company's support team exports 100,000 resolved Zendesk tickets, keeps those with high CSAT scores and verified resolutions, and strips customer names, emails and account IDs with its own scripts before saving them as a JSONL dataset. The dataset is uploaded to Ertas Studio, and two senior agents use Data Craft to correct and approve a few hundred examples for the trickiest billing and account issues. In Studio, the team fine-tunes Llama 3.1 8B with LoRA on the company's troubleshooting and account management domains and runs three configurations in parallel. They compare each run against a held-out set of 5,000 tickets, scoring its answers against the resolutions agents actually sent and against the answers from their current RAG chatbot. The best model is exported from Hub, served with vLLM on the company's own GPU server, and connected to a Zendesk webhook. It drafts replies to tier-1 tickets, and the team decides which categories to automate fully based on its own review of those drafts, sending the rest to agents for approval.

Compliance & Security

Support conversations often contain customer PII such as names, emails and account details. Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas: remove or mask PII with your own tooling before upload, and delete datasets from Studio when the project ends. After export, the model can run on your own infrastructure, so customer messages sent for inference stay there. GDPR, CCPA and contractual privacy obligations remain your organization's own work, and the training set should go through your usual privacy review.