"Pet-friendly" is a LIE, its the most meaningless tag in hospitality. Nobody verifies what it means and its hairy out there, I see breed restrictions, weight limits, off leash space, fee structures, and what made me crazy is that none of it standardized, none of it checked. So I built a structured dataset covering 58,000 venues and 85,000 trails across 56 countries, all verified, the validation and data decay pipeline at scale is the hard part.
Kali is the interface.
She is a WhatsApp native AI concierge. Dog owners message her, she answers from real data, not scraped tags. She can tell them where the best dog friendly bars, restraurants, dog parks, trails, events, and activities are in the 50+ countries but especially the US and UK.
Architecture:
Cloudflare Worker (TypeScript), D1 for conversation state, KV for user profiles
Claude Haiku for all conversations. p95 latency 9 seconds.
Cost per conversation roughly $0.01-0.03
Vectorize semantic search across venue and trail data. Speaks 30+ languages natively, not translated. Same prompt, Haiku just handles it.
The Opus supervisor (I call it SELFIMPLOOP):
This is the bit I want to talk about. Haiku is fast and cheap but it drifts. It gets overconfident. So every day, Opus reviews every Haiku conversation from production, grades them against failure categories (fabrication, rule violations, false confidence, missed searches, tone failures), and proposes prompt rewrites. Nothing auto applies, I still review every change before it goes live but jesus its doing a better job than I am.
A few days ago this system caught something serious. An dog owner told Kali what her dog was sick and she needed an emergency vet, Kali made me proud and pulled two emergency 24 hour vets in the users city from vector memory in seconds, I won that user for life.
What is scaring me is that the referral flywheel works too well. Users bring in other users without being asked and am scared I have built something bigger than my hardware. WhatsApp is the right channel, people are sick of apps, gated UI's and general knowledge AI, they want what they want in their hand where they already in conversation. Thats just the way it is now I think.
The 24-hour WhatsApp session window is annoying but manageable, if it expires (which it inevitably does), we fall back to email. Running Haiku at scale is absurdly cheap but I would rather have Opus on every conversation, I will when I can am not sure why dog owners should have to suffer on Haiku.
What I have not figured out:
Converting tire-kickers to repeat users, most people try it once and the ones who come back are gold. Data is thin in lots of countries, am working on it. The mini site is kaliconcierge.com (currently in ten languages). Happy to chat with you about this, she is my baby and I want to make her strong, am a big fan of hostile audits so bring it.
I think I should turn off my refferal flywheel while I get a handle on what I have right? Or NO because I want lots of users who might destroy me but maybe not?
Show HN: The cutest WhatsApp concierge for dog friendly travel · HackerTrans
Kali is the interface.
She is a WhatsApp native AI concierge. Dog owners message her, she answers from real data, not scraped tags. She can tell them where the best dog friendly bars, restraurants, dog parks, trails, events, and activities are in the 50+ countries but especially the US and UK.
Architecture:
Cloudflare Worker (TypeScript), D1 for conversation state, KV for user profiles Claude Haiku for all conversations. p95 latency 9 seconds.
Cost per conversation roughly $0.01-0.03
Vectorize semantic search across venue and trail data. Speaks 30+ languages natively, not translated. Same prompt, Haiku just handles it.
The Opus supervisor (I call it SELFIMPLOOP):
This is the bit I want to talk about. Haiku is fast and cheap but it drifts. It gets overconfident. So every day, Opus reviews every Haiku conversation from production, grades them against failure categories (fabrication, rule violations, false confidence, missed searches, tone failures), and proposes prompt rewrites. Nothing auto applies, I still review every change before it goes live but jesus its doing a better job than I am.
A few days ago this system caught something serious. An dog owner told Kali what her dog was sick and she needed an emergency vet, Kali made me proud and pulled two emergency 24 hour vets in the users city from vector memory in seconds, I won that user for life.
What is scaring me is that the referral flywheel works too well. Users bring in other users without being asked and am scared I have built something bigger than my hardware. WhatsApp is the right channel, people are sick of apps, gated UI's and general knowledge AI, they want what they want in their hand where they already in conversation. Thats just the way it is now I think.
The 24-hour WhatsApp session window is annoying but manageable, if it expires (which it inevitably does), we fall back to email. Running Haiku at scale is absurdly cheap but I would rather have Opus on every conversation, I will when I can am not sure why dog owners should have to suffer on Haiku.
What I have not figured out:
Converting tire-kickers to repeat users, most people try it once and the ones who come back are gold. Data is thin in lots of countries, am working on it. The mini site is kaliconcierge.com (currently in ten languages). Happy to chat with you about this, she is my baby and I want to make her strong, am a big fan of hostile audits so bring it.