Fishnerd is a full-stack fishing companion for Georgia’s freshwater fish, built around open-weight AI for photo-based identification. Its Python backend runs the pretrained BioCLIP model locally, using image and text embeddings to rank potential matches across a catalog of 290 species and other taxonomic entries. This approach requires no paid inference API or custom model training and lets developers control model hosting, hardware, and inference without sending identification photos to a third-party AI service. Predictions remain suggestions that users review and confirm. Around this identification workflow, Fishnerd provides a personal catch journal, optional photos and measurements, kept-or-released tracking, and a collectible Fishdex whose illustrated fish unlock as users discover them. Personal records, quests, XP, and community leaderboards turn catch logging into an ongoing exploration experience.

The repository combines an Expo and React Native frontend written in TypeScript with a FastAPI backend and persistent SQLite storage. Its open-weight inference pipeline supports configurable model selection and CPU or supported GPU execution, reuses cached model weights and species embeddings, and keeps identification separate from catch storage and regulation checks. Users can also view sourced Georgia statewide guidance for size limits, daily allowances, and seasons, with retained catches contributing to daily-limit tracking. Supporting scripts and automated tests cover model execution, catalog and regulation coverage, catch handling, collection progress, quests, and artwork. Fishnerd demonstrates how openly available model weights can power a specialized application without a proprietary inference dependency, while acknowledging that identification accuracy remains unvalidated and statewide guidance does not determine whether a fish may legally be kept.