Human rights measurement has long been constrained by a familiar trade-off: existing indicators are either broad and infrequent, or fine-grained but painstakingly hand-coded and difficult to replicate. This paper shows how large language models can break that trade-off. I develop an LLM-based extraction pipeline that converts unstructured reports from local Turkish human rights NGOs into a structured, daily-level panel, yielding roughly 30,000 unique violation records spanning 2013 to 2023, at a temporal and geographic resolution existing cross-national datasets cannot match. Beyond the Turkey case, the contribution is methodological: a scalable, replicable framework for turning NGO reporting, a vast and underused textual archive, into systematic human rights data, with direct extensions to other countries and rights domains. The paper reports validation results, discusses failure modes and bias checks for LLM-based coding of sensitive political content, and situates the approach within the conference's broader agenda on automated data curation and LLM applications in policy research.