A Hot Pot, a Language Barrier, and One Minute of Setup
The Hot Pot Dinner
We were in Vancouver for Web Summit, staying at an Airbnb the way cash-strapped founders do: shared room, occupied house, just grateful for the roof over our heads.
Our host was a Chinese woman who grew most of her own food and spoke no English. She communicated through her son when he was around, other bilingual guests, and a translation app when no one else was available. We communicated with nods and hands. It worked, mostly. But "mostly" has a ceiling.
On our first night, she made us hot pot. We sat down with the other guests and she was chatting up a storm, and Yash and I were so lost! She started using her translation app, but the output left something to be desired. We'd have to decipher the English to figure out what she probably meant. Other guests would help, but we felt guilty disrupting their dinner for them to be the middleman.
Between mouthfuls, I opened Mettara on my phone and created a new AI persona. The instructions were about four sentences: respond only in Chinese, translate whatever I type, keep the phrasing natural. That was it. No fine-tuning, no integrations, no configuration beyond that.
It's worth noting that I'd tried the same thing with Claude first. I spent several minutes working on it. It didn't hold. Claude isn't designed to take in English and output only a clean Chinese translation and nothing else. It kept adding commentary, context, alternatives. Useful in other situations. Wrong tool for this one. One of the bilingual guests confirmed our first Mettara output was accurate.
For the rest of the evening, she used her app. I used ours. Back and forth.
She told us about the career she left in China, where she ran her own food production factory, overseeing nearly 500 people at its peak. She talked about why she'd moved, what she was building here, how she thought about the future. She asked us what we were working on and had opinions about it. Real opinions, not polite ones.
At some point she said, paraphrased through two apps, that she thought AI was one of the most important things to happen in her lifetime. Not because of efficiency or automation. Because it let her talk to people she couldn't talk to before. It made the world smaller and individuals more capable.
What This is Actually About
We didn't build Mettara for dinner parties. We built it for B2B SaaS products that need AI embedded inside them. They need operators that understand context, respect tenancy boundaries, and can collaborate with other AI agents and humans in the same workflow.
But that evening demonstrated something we think about a lot: the value of a purpose-built AI persona isn't the model underneath. It's the specificity of the role.
A general-purpose AI assistant does many things adequately. An AI with one job, like "translate what I type into natural Chinese" does that one thing exactly right, and it's trustworthy because of that narrowness. You know what it will do. You know what it won't.
That same principle applies inside a product. A billing AI that only touches billing data. A customer success AI that only surfaces information relevant to that account. Narrow scope isn't a limitation; it's what makes the output trustworthy enough to act on.
The hot pot conversation worked because both parties trusted the translation. She trusted hers. I trusted mine. Neither of us was second-guessing the output mid-sentence. That's the experience we're trying to give enterprise users with every AI operator we help teams build.
The Setup Time
Literally one minute. That's it. I gave a quick description to our AI designed specifically for creating other AIs. From "I wonder if this would work" to first confirmed translation.
That's not a product demo. But it is a data point about what it means to create a purpose-specific AI that earns trust fast.
If you're building something where that kind of speed-to-specificity matters, we'd like to talk.