The Kryotta blog
Guides and comparisons on working with every AI model — Claude, Gemini, Llama, Mistral, DeepSeek and more — plus what we're building.
I used Compare to test 4 AI models on 85 Paystack dispute emails: which one won per claim
I tested Claude, Gemini, Llama, and GPT-OSS on 85 real Paystack dispute emails over two weeks. Here's which model won the most cases and why.
I used AI for 40 restaurant order queries during load-shedding: checklist for what works
A Cape Town restaurant owner tested Claude and Gemini on 40 real WhatsApp order queries during load-shedding. Here's what AI handled alone, what it botched, and the cost per response.
I used Compare vs Auto routing for 180 Shopify returns after Boxing Day: which approach per query type
I tested Kryotta's Compare and Auto routing features on 180 Boxing Day return requests. One saved nine hours. The other caught edge cases the system missed. Here's which approach worked best for each query type.
I made 48 product shots with Stability AI then real photos: which got M-Pesa orders
I tested AI-generated product images against real photos across 12 products in my Nairobi accessories business. One approach hit 4.1% conversion; another got Instagram flagged twice.
I used 5 prompts to explain hallucinations and grounding to my WhatsApp vendor team: which one they actually understood
I tested five ways to explain AI hallucinations to my WhatsApp vendor team. Only one actually stopped them from trusting false information—and it wasn't the technical explanation.
I used 4 AI models to write 120 Jumia product descriptions: which one got clicks in Accra
A Jumia seller tested Claude, Gemini, Llama, and Mistral on 120 product descriptions in Accra. Claude won on clicks with mobile-first copy that converted 3.4%—here's what each model cost and how shoppers responded.
I used AI to answer 6 questions about Black Friday prep that freelancers actually ask
I tested Claude, Gemini, and DeepSeek to answer the questions freelancers actually ask about Black Friday prep—from email sequences to product images that won't get rejected by Meta.
I made 38 product shots with Stability AI then TikTok Shop ads with Luma: checklist for what fails
I generated 38 product images with Stability AI for a TikTok Shop launch, then half failed review. Here's the checklist that separates what platforms approve from what customers actually trust.
I used 3 AI models to write 95 Klarna late-payment reminder emails: which tone got paid
I tested three AI models writing 95 overdue payment reminders in different tones. The results showed tone mattered far more than which model I used—and the most effective approach surprised me.
I used 4 AI models to write 160 SEPA mandate rejection emails: which one per bank
A Berlin bookkeeper tested four AI models on 160 failed SEPA mandate emails and picked different ones for each bank type. Here's which model worked best for Sparkasse, N26, Revolut, and others.
5 myths about new AI models that waste your USD budget on WhatsApp commerce
Newer AI models don't always justify their cost for WhatsApp shops. We tested real customer messages and found that expensive reasoning models often waste money on simple queries like stock checks and delivery questions.
I made 42 product shots for Boxing Day sales with Stability AI: which prompts passed and which got returns
I generated 42 product images with Stability AI for Boxing Day sales. Twenty-six went live, but some triggered returns. Here's which prompts worked, which failed, and what cost me £340.