The engineering blog
Where a language model helps and where a person stays in charge: content, tooling and the guard rails that make AI safe to use.
Many tasks labeled agentic don't need an agent. Four questions, built on Gartner's advice and the model makers' own guidance, show whether yours might. Six more tell you what to ask whoever builds it.
Bozhidar Ivanov
Uber used up a year's AI budget in four months, then steadied its AI spending with cheaper models for some tasks, caching, visibility and caps. Here is how the same controls keep an AI feature's running cost predictable.
Bozhidar Ivanov
An AI-built app can work perfectly and still let anyone read or change its data. Before launch, check the access rules on every table, keep secret keys off the page and run your platform's own security checks.
Bozhidar Ivanov
Two OpenAI requests share long instructions and a cache key, yet the second reports zero cached tokens. See where the cache breakpoint sits, why a changing value such as the current time blocks reuse, and how to verify a fix.
Ivaylo Valchev
TypeSafe's Jev returns typed decisions instead of text. What it does, how open alternatives like Laya, Kev and OpenJev compare, and how to test one on your own data.
Bozhidar Ivanov
OpenAI's dots are always-on AI agents with their own cloud computers. Who can use them, how their safeguards work and what they mean for your business software.
Bozhidar Ivanov
An LLM that writes into an existing database uses the document's names, not yours, and "not found" quietly turns into "new". That is how one import created 790 duplicate cars. The fix was not a smarter model. It was teaching the pipeline the database's vocabulary, and repairing the data from decisions written down in advance.
Bozhidar Ivanov
llms.txt is a proposed file that points AI tools at a site's key pages. What it does, what it doesn't, and how to keep one from going stale.
Bozhidar Ivanov
Arc AI lets a language model convert articles into CMS-ready fields, never rewrite them or decide what gets published. AI output passes the same checks as a human editor's input.
Bozhidar Ivanov
An AI voice agent can stop when interrupted and then resume the old answer. This case shows how to test the symptom, identify where the audio came from, and verify that the next reply still works.
ArcAI Team
