How Xevra works under the hood

Inside Xevra there’s no neural network making up names. There’s a deterministic engine: morphology, phonetics, and eight measurable quality criteria — no "making things up" in the sense a language model does it.

The path of one request

The business description is first broken into keywords and mapped to semantic clusters — roughly, the engine understands that "neighborhood coffee shop" connects to the themes "coffee," "cozy," "morning." From those themes a pool of semantic roots is assembled, and five different morphology strategies build candidates from it: sometimes a root plus a suffix, sometimes a fusion of two roots, sometimes a light phonetic mutation. Every candidate passes a filter: disallowed sound combinations, profanity, matching a real word of the language, matching an existing brand. Whatever survives the filter is scored on eight parameters and lands in the final list.

Why this matters, not just a technical detail

• Reproducibility. The same request and the same run parameter give the same result — you can come back to the list a week later without losing a name you liked.
• No name ships without an explanation. For five of the six generation strategies, you can always see the root a name grew from — not "the model decided so," but a concrete semantic link you can verify with your own eyes.
• Zero LLM cost for the generation itself. There is a language model in the system, but it only kicks in at the last step — to explain and flag risks on an already-finished top list. It never invents names itself — a deliberate architectural decision, not something we just haven’t gotten around to.

What the engine doesn’t do

It doesn’t guarantee you’ll personally like a name — taste isn’t measured by an algorithm. And it doesn’t replace a trademark lawyer’s review in your specific category — the filter screens out known brands and obvious conflicts, but it isn’t a substitute for full legal due diligence before registering a business.