The Sunday Letterhttps://blog.tarkika.com/ One essay every Sunday on building AI that decides, with the code. By Sridhar Mukkandi.en The Jevons paradox of decisionshttps://blog.tarkika.com/the-jevons-paradox-of-decisions/https://blog.tarkika.com/the-jevons-paradox-of-decisions/Sun, 04 Oct 2026 07:00:00 +0000DecisionsIn 1865 Jevons saw better engines make Britain burn more coal, not less. Cheap AI decisions follow the same pattern, with a twist that lands on the people who review them. Act, review, or escalate: agents that know their limitshttps://blog.tarkika.com/agents-that-know-their-limits/https://blog.tarkika.com/agents-that-know-their-limits/Sun, 27 Sep 2026 07:00:00 +0000AgentsA good model isn't a working decision system. The team you have, what's at stake, and what happens on the day the model is down decide whether an agent can safely act on its own. Six ways to make one decisionhttps://blog.tarkika.com/six-ways-to-make-one-decision/https://blog.tarkika.com/six-ways-to-make-one-decision/Sun, 20 Sep 2026 07:00:00 +0000DecisionsRules, a trained model, a text classifier, an LLM writing JSON, an LLM judge, and a model built to decide. One dataset, one fair contest, and a scorecard with every column that matters. Your LLM's 0.95 isn't a probabilityhttps://blog.tarkika.com/your-llms-0-95-isnt-a-probability/https://blog.tarkika.com/your-llms-0-95-isnt-a-probability/Sun, 13 Sep 2026 07:00:00 +0000LLMsAsk a language model for JSON with a confidence field and you get a tidy number back. Before you put a threshold on it, look at what that number is made of. When 0.8 should really mean 80%https://blog.tarkika.com/when-0-8-should-really-mean-80/https://blog.tarkika.com/when-0-8-should-really-mean-80/Sun, 06 Sep 2026 07:00:00 +0000CalibrationA model can rank every alert in the right order and still lie about how sure it is. Here is how to catch it with nothing but grouping and counting, and the fixes that work. Why 0.5 is almost always the wrong thresholdhttps://blog.tarkika.com/why-0-5-is-the-wrong-threshold/https://blog.tarkika.com/why-0-5-is-the-wrong-threshold/Sun, 30 Aug 2026 07:00:00 +0000DecisionsEvery tutorial turns a probability into an action with one line at 0.5. On a real security queue, that line quietly lets thirty-seven threats a day through. Here is where the lines should go.