Ertas for E-Commerce

The Challenge

E-commerce businesses operate at a scale where even small improvements in product discovery, description quality, or customer understanding translate directly into revenue. AI can power better search, personalized recommendations, automated product descriptions, and conversational shopping assistants, but generic models have no knowledge of your specific catalog, taxonomy, or customer base. They produce generic product descriptions that sound like every other retailer, misinterpret niche product queries, and fail to capture the brand voice that differentiates your store.

Cost is the other critical factor. At e-commerce scale (millions of product pages, thousands of customer queries per minute, and real-time recommendation requests on every page load), per-token pricing from third-party API providers becomes prohibitively expensive. A large catalog retailer generating AI product descriptions for 500,000 SKUs or serving a conversational shopping assistant to millions of visitors faces API bills that dwarf the value the AI creates. The economics only work if inference costs are predictable and decoupled from per-request pricing.

The Solution

Ertas lets e-commerce teams train AI models on their own product domain and run them without a per-token API bill. In Ertas Studio, merchandising and data teams upload JSONL datasets built from their product catalog, customer reviews, search query logs and brand style guides, or build examples in Data Craft, Studio's guided dataset builder, where the people who know the brand review and approve each one. Training runs on Ertas's cloud GPUs through a visual canvas, and several runs can go in parallel so the team can compare them before choosing one. The resulting models write product descriptions in your brand voice, read ambiguous shopper queries with catalog-specific context, and classify products against your own taxonomy.

Each trained model appears in Hub, where you download it as GGUF for llama.cpp or Ollama, or as full safetensors for vLLM and other server runtimes. Run it on your own servers and inference costs track the hardware you run, so they stay predictable as query volume grows, and you can add capacity ahead of peak shopping events. Train separate models for separate jobs (one for product descriptions, one for search query understanding, one for customer service) and keep each one as an asset your team owns. For teams that would prefer a managed service, the adaptive model endpoint moves a growing share of one task's requests from the frontier model you use today to a custom model trained on your own requests, once that model passes your evals.

Key Features

Example Workflow

A direct-to-consumer fashion retailer with 80,000 SKUs wants to automate product description generation and improve on-site search relevance. The merchandising team exports product attributes, existing descriptions, de-identified customer reviews and search query logs as JSONL datasets and uploads them to Ertas Studio, then uses Data Craft to review a sample of its best descriptions with the brand team. In Studio, the team fine-tunes Mistral 7B twice: one model writes brand-voice product descriptions from raw attributes, the other maps natural-language search queries to product categories. They run a few configurations in parallel and compare them against a held-out set of products the brand team has already written copy for. Both models are exported from Hub and served with vLLM on the retailer's own GPU servers. The description model drafts copy for 20,000 new seasonal products in an overnight batch job, and the brand team spot-checks the drafts against its tone guide before they go live. The search model interprets queries like 'breathable summer office pants' as a match for the 'lightweight trousers' category. Inference costs now track the servers the retailer runs, so heavier query volume no longer raises a per-token bill.