What is kira-project-site?

The Kira Project is a personal engineering blog built around a homemade AI system called Kira. Kira runs entirely on one Apple Silicon computer, using a single 9 billion parameter model to handle every role in its pipeline: routing requests, acting as specialists, synthesizing answers, and planning multi step tasks. It calls on roughly 50 tools while working, and none of that work is sent to cloud servers, everything runs locally on that one machine. The project is built and maintained by one person, working alongside AI tools as a kind of pair programming partner. Decisions about the system's design are written down as formal records, so choices about how Kira works are documented rather than left to memory. To check that the system behaves correctly, the team runs more than 1,100 live tests that run against the actual system and confirm it does what it is supposed to. This repository is not the code for Kira itself. It is a publishing site, a place to share posts about what the team has measured while building and running Kira. One post describes a way of recording what the AI could not verify or trust while doing its job, creating a tamper resistant log that would visibly break if someone tried to edit it after the fact. A patent application has been filed covering this idea. Another post explains a performance problem the team found, where the model was spending close to half its processing time reprocessing text it had already seen, and how fixing that cut total processing time by roughly a third to nearly half. That fix is also the subject of a separate pending patent. The site's writing and images are shared under a Creative Commons Attribution license, meaning others can reuse them as long as they give credit, and the authors invite readers to replicate their measurements and share what they find.