A customer searched for “inline skates size 42” on one of our Merchant's Shopify stores.
Our AI recommender suggested t-shirts.
I spent 2 hours debugging embeddings, cosine similarity, and the entire pipeline. Turns out we were feeding it raw product HTML. So “inline” didn’t mean inline skates. It meant inline-flex, inline-block, etc. The AI wasn’t broken. It was just ranking the products with the most CSS.
4-line fix: strip HTML before embedding.
Most AI bugs are just data bugs with better branding.
I build AI customer service agents for a living, and I've been watching this week's meltdowns about chatbot loops with a guilty conscience.
MakeMyTrip, Capital One, Zomato, Air Canada. Customers stuck in loops where no human ever shows up. I get why people are furious, because the entire industry designed it that way on purpose. The default goal behind almost every AI support tool is to deflect the ticket, close it before a person ever has to touch it. "How many conversations did the bot handle without a human" is literally the sales pitch.
I went the opposite direction with Nousu. Our agent checks if a human is online in the helpdesk. If nobody's there, it captures the customer's info, reassures them someone's on it, and sends them an email if things take too long. If a human is online but slow to respond, the agent nudges them again and messages the customer that it'll be a bit longer but they're not forgotten.
The whole point is that it sits next to the human like a colleague, doing the work a good teammate would do: covering when you're away, keeping the customer calm, pulling you back in the moment you're needed.
Building support agents that actually fix your shit instead of just deflecting it feels like a completely different product, and honestly I'm not sure the industry even wants to go there yet.
Our https://t.co/8JIiaHGPV5 support agent just did something I didn't build it to do.
I've been experimenting with @PhotonHQ to let the agents feel more like a colleague through iMessage. This week it found a broken flow, reproduced it, took a screenshot, messaged me with exactly what went wrong, and contacted the customer.
Now you'll just get a message from the agent with the bug, the screenshot as proof, and the exact steps to reproduce it.
Building support agents that actually fix your shit instead of just apologizing for it feels like a completely different product.
@Dessn_ai Looks good! Does it 'solve' designing with AI better than Claude Design and such? Most AI design tools still create generic layouts and don't feel polished.
We spent two years building an AI that takes one goal all the way to the inbox.
That required inventing a new email language. One that renders perfectly in Gmail, Apple Mail, dark mode, even Outlook 2003.
Introducing @Migma_AI
It designs one campaign and adapts it for different audiences and languages. You approve it, and it sends.
It also sends from your own domain, learns from every send, and connects to Cursor, Claude Code or Codex so your agent can wire up your app's emails.
Migma Pro is free for 7 days and 50% off forever for first 200 supporters.
Repost and comment "Migma" and we'll send the access over.
Live today ↓
@levelsio@mitsuhiko I remember sleeping with a fan pointed directly at my face, because of course proper air conditioning is basically unheard of in the Netherlands. Europe’s climate control standards really are non-existant. 🙃