Building Modern Tools for Computational Science in Flutter to Track the American Chestnut

Abstract

The American Chestnut is a functionally extinct hardwood tree that once accounted for 10-25% of trees in forests in the eastern United States, primarily along the Appalachian range. The population was largely eliminated due to a blight inadvertently introduced from nursery stock of two species of Asian chestnut trees imported for orchards. Efforts to restore the American Chestnut rely on data gathered from decades of selective breeding in research orchards by professional foresters and observations of the wild, remaining population provided by volunteers. Tracking wild American Chestnuts can be tricky due to terrain, network access, and data collection methods, and it is further complicated by the need to use mobile devices in the woods and workstations in the lab. We present a suite of tools in a codebase called Stager, built with Flutter and used to power a cross-platform, multi-tenant, geographical information system specifically designed to track these trees wherever they may be found. This application, Tracee, combines the tools in Stager to provide a mobile "field tool" app that merges live mapping and geolocation with voice models and mobile cameras to pinpoint trees while capturing images and transcribing live phenotyping notes in a forester-friendly way. Data from the field tool syncs with desktop and web versions of Tracee to put the field data directly in the hands of research scientists. This work presents the end-to-end process of creating and releasing both products, which heavily relied on large-language models (LLMs) in every part of the software development lifecycle, including business modeling and design. The specific challenges of utilizing AI-based code generation are discussed with a special comparison to traditional code generation techniques, such as deterministic project template-based generators. The single most important takeaway for the audience will be the presentation of a straightforward method for driving LLMs to successful coding goal completion for developing these kinds of tools will be presented to fill in the practical gaps. The combination of these techniques to create new scientific applications and tools in less than a year with only one developer demonstrates the type of impact that AI-enabled engineering can have when used correctly, with a focus on empowering instead of replacing research software engineers.