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 3 AI models to review 60 pull requests: checklist for which model per task
I tested Claude, Gemini, and DeepSeek on 60 real pull requests. Each model excelled at different tasks—here's the checklist for which one to use when.
I used AI to reply to 150 Etsy messages during Black Friday: which model kept up
I automated 60% of my Etsy customer messages during Black Friday using AI routing. Here's how I set it up, what failed, and which models actually kept up with 150 messages.
Claude vs Gemini for M-Pesa statement reconciliation: I tested 80 till reports
I tested Claude and Gemini on 80 M-Pesa till statements to see which AI model could cut my three-hour daily reconciliation down. Here's what actually worked.
I made 8 product videos for Flipkart ads with AI — which ones got clicks and which got skipped
I tested three AI video tools on eight Flipkart product ads with a ₹5,000 budget. Four videos hit 4%+ click rates. Four flopped. The difference wasn't the model—it was what I asked it to do.
I made 40 product shots for Flipkart and Meesho with Stability AI: which prompts got clicks
I tested 40 AI-generated product shots across Flipkart and Meesho to find which prompts actually drive clicks. Clean backgrounds won on Flipkart; lifestyle shots dominated on Meesho. Here's the data.
I used 3 AI models to debug a Stripe webhook timeout: which reasoning model found it
A Black Friday payment crisis: orders processed but never delivered. Three AI models tackled the same logs. Only one found the Nginx timeout mismatch killing Shopify API calls.
I used AI to write 50 restaurant menus and WhatsApp order replies — which model kept the Ghanaian English natural
I tested Claude, Gemini, and Llama on 50 real food-stall tasks—writing menus and WhatsApp replies. One kept the warmth and rhythm of Ghanaian English. The others sounded like call centers.
5 myths about AI that waste your money: tokens, hallucinations and context windows explained
You don't pay per question—you pay per token. Context windows aren't memory. And hallucinations aren't lies. Here's what actually costs you money with AI.
I used 3 AI models to review 25 TypeScript pull requests: which one caught the VAT rounding bug
I tested Claude, DeepSeek, and Gemini on 25 TypeScript pull requests. DeepSeek caught a VAT rounding bug that would have cost a client thousands—here's how each model performed.
I made 15 TikTok Shop product videos with AI — which ones got sales and which got scrolled past
I generated 15 AI product videos for my TikTok Shop in two days. Four converted. Here's what worked, what flopped, and why lighting mattered more than motion.
I made 25 Jumia product photos with Stability AI — which prompts got clicks and which got returns
I tested Stability AI on 25 phone case listings. Thirteen worked. Twelve failed badly—customers returned items saying they looked nothing like the photos. Here's what the prompts got wrong.
Claude vs Gemini for GST invoices: I tested 100 B2B bills and refund notes
I tested Claude and Gemini on 100 real GST invoices, credit notes, and inter-state bills. Here's which model got the tax splits right, which hallucinated HSN codes, and the prompts I now use weekly.