DOMAR — product-description agent
An online door & window shop needed unique product descriptions, with no invented product facts.
The agent collects the manufacturer's catalog text, a local LLM (Qwen 3.5 on llama-server) writes only from that source, then gates run from cheapest to most expensive: rule checks for invented facts, a similarity gate for uniqueness, and an independent Gemini-based reviewer agent with a local fallback.
Failure mode is “no description”, never “wrong description”. Measured end-to-end runs on real catalog items (acceptance rate, tokens/s, latency) go to the owner as the basis for choosing models.