AIVUKAT

A general-purpose chat model can answer a legal question fluently but without grounds. Every aivukat answer had to be tied to a statute article or a court decision, and to say so when it could not be. Case law runs to millions of decisions, and the questions come from lawyers and from citizens unfamiliar with legal language alike. On top of that, the questions had to be answered on a single server without a graphics card, never sent to an external AI service. There was no brand, no data and no infrastructure; all of it had to be built from scratch.

Brand: the name, logo, colours, typefaces, tone of voice and brand guide were designed from scratch. Data: legislation is collected from mevzuat.gov.tr with a collector we wrote; case law is selected from an open dataset of Yargıtay, Danıştay, UYAP Emsal and Constitutional Court decisions, with duplicate decisions removed. Retrieval: texts are split into passages and turned into meaning vectors with a multilingual model, and an index that scales to large data is built on SQLite and FAISS; it is produced locally on a GPU and moved to the server. The answer model was chosen through a 40-question evaluation: Gemma 4 E4B, which gave the best grounding. It runs on a single server with no graphics card, on an accelerated build of llama.cpp; the API, the model, the site and TLS share one machine, and the setup can be repeated with a single command. Answers come with numbered sources and their citations, and tapping a source opens its text alongside; without enough sources, no answer is written. The interface is a single-screen chat: users choose between a "Vatandaş" (citizen) and an "Avukat" (lawyer) style of explanation, chat history stays only in the user's browser, and every screen states that the site gives no legal advice. A launch set of Instagram posts and short videos was produced as well.