Show HN: Using sound symbolism and multi-agent AI to generate brand names

I built an AI naming tool that applies psycholinguistic research to brand name generation. The interesting part isn't that it uses AI — it's how the agents are structured and what they're optimized for.

The core problem: if you ask any LLM to name a business, you get the same [Adjective][Noun] compounds. NovaTech. BrightPath. SwiftFlow. They're linguistically dead — no phonetic texture, no semantic depth, high cognitive fluency but zero distinctiveness.

The pipeline has six stages:

1. A discovery agent analyzes the business and produces a strategic brief. Critically, it also generates a "tangential category" (something completely unrelated, like "a luxury candle brand" for a SaaS tool) and a "disguised context" (an adjacent industry).

2. Three creative agents run in parallel, each with a different framing of the same brief. One works honestly from the brief. One is told it's naming the disguised context. One is told it's naming the tangential category. The disguised and tangential agents consistently produce more interesting names because they're freed from category conventions — the LLM can't fall back on the obvious industry vocabulary.

3. A linguistic filter scores all ~90 candidates using sound symbolism research: - The bouba/kiki effect (round sounds like b, m, l, o map to friendly/soft; sharp sounds like k, t, p, i map to edgy/precise) - Processing fluency (ease of pronunciation, spelling, recall) - The Von Restorff isolation effect (distinctiveness from category norms) - Consonant/vowel balance and syllable structure

   Each name gets a 0-100 score. Top 25 survive.

4. Domain availability across ~280 combinations (7 TLDs x multiple variations).

5. A synthesis agent ranks the final 10. This stage uses Claude instead of OpenAI — the ranking requires balancing semantic relevance, brand fit, sound symbolism scores, domain availability, and "polarization potential" (names that provoke a reaction tend to be stronger brands). Claude handles this kind of multi-factor holistic judgment noticeably better in my testing.

6. Trademark screening against the USPTO database, cross-referenced with the Nice classification classes identified in stage 1.

The two-model split was a pragmatic choice. GPT-4o-mini is fast and cheap for structured generation and analysis (stages 1-4). Claude Opus is better at the subjective ranking tradeoffs in stage 5 but would be too expensive to run across all the parallel creative agents.

The linguistic scoring is the part I find most interesting. Sound symbolism is well-established in psycholinguistics but rarely applied systematically to naming. Lexicon Branding (who named Sonos, Pentium, Blackberry) uses these principles — the "s" sounds in Sonos evoke smoothness and flow, which maps to their product experience. The tool tries to do the same analysis programmatically.

Genuinely curious what HN thinks of the names it generates. Try it with a business you know well and see if the output feels different from what ChatGPT gives you.