The volume of scientific publications has grown rapidly in recent years, making it increasingly difficult for researchers to maintain an overview of the state of the art. For example, a Web of Science search for `hydrogen production' alone returns over 100,000 results. When building a computational model of a hydrogen value chain, identifying which mathematical models and methods from the literature are best suited for the task is a substantial effort. Similar challenges arise with patent databases and in-house, proprietary document collections — and in such cases, privacy may preclude cloud services. To address this, we introduce Raven, an open-source AI system for navigating large scientific literature collections. Raven is a 100% local, privacy-first tool that runs entirely on a single workstation with an NVIDIA GPU. The system uses open-weight, pretrained AI and ML models throughout. Raven consists of two core applications. First, Visualizer performs topic discovery on a BibTeX citation database. It computes semantic embeddings of titles and abstracts, clusters similar entries with HDBSCAN, and projects the result into a 2D map with t-SNE. This enables rapid, interactive exploration of the topical structure of a large literature collection. Secondly, Librarian is a conversational search and analysis tool using a local large language model (LLM). It employs hybrid keyword-semantic retrieval-augmented generation (RAG) to ground the LLM's answers in the dataset. We demonstrate Raven on a dataset containing the metadata of approximately 12,000 hydrogen production studies from Web of Science. Topic analysis with Visualizer reveals major research clusters including photocatalysis, biological fermentation, and steam reforming. From a computational mechanics perspective, numerical modeling studies in this dataset are few, and they appear scattered across all topic areas rather than forming a distinct cluster. We then use Librarian to explore individual areas in more detail, with an eye toward subprocesses in a hydrogen value chain. As an example, we consider photovoltaic hydrogen production. Given variable solar irradiance, what modeling approaches are relevant --- for example, for cell efficiency, electrolyzer dynamics, or storage? Raven is open-source software, available at. As locally hostable open-weight LLMs continue to improve, we expect the utility of local AI tools for scientific work to increase correspondingly.