Goal
This week, I wanted to start developing an independent project and explore Terrier’s capabilities. My first idea connects back to my interest in birding and birdwatching.
Workflow
I first started by looking online for the theory behind weather-driven bird migrations. This was key to understanding which Terrier data would be most useful or valuable to my project. I then needed to evaluate each one based on its level of influence on bird migration to contribute to the overall score of birdwatching success for an area.
Using this list of tiered inputs, I created a dashboard in Codex that calculated a bird score for each station based on modeled migration patterns. I also want to integrate the more species-focused element of birding. I added a layer that pulls data from iNaturalist and eBird to display birds reported recently within 25 miles of the selected station.
Prompts Used
Codex Prompt: Build a React + TypeScript dashboard prototype that uses Wet Dog Weather Terrier and MapLibre to visualize radar and weather data alongside a simple migration probability model. The prototype should estimate the likelihood that radar echoes represent nocturnal bird migration versus precipitation using synthetic data, display probability overlays on a map, provide forecast-based outlooks, and explain the factors influencing each probability score. Terrier should be used only for weather visualization, not migration detection.
What Worked
AI was valuable in translating bird-migration forecasting concepts into a usable prototype: converting wind direction, precipitation, storm risk, ceilings, visibility, pressure trend, frontal passage, timing, and low-level jet strength into a structured scoring model; building the Terrier/MapLibre layer controls; and connecting those conditions to be easily digestible.
What Didn’t Work
AI can build and revise this dashboard quickly, but it struggles with judging visual map behavior and scientific calibration without direct testing. That led to repeated marker and layer-order fixes: compilation and HTTP checks passed, but they did not prove the map looked or behaved correctly. The remedy is browser-based regression testing for zoom, pan, layer toggles, marker selection, and mobile layouts.
Likewise, the migration score uses scientifically relevant inputs but is not a validated forecast model. It should be presented as a prototype suitability until calibrated against radar migration data, BirdCast outputs, and expert review.
Lessons Learned
For similar AI-assisted scientific dashboards, AI is most valuable as a rapid prototyping and systems-integration tool, helping translate domain knowledge into data structures, scoring logic, interactive maps, and clear interfaces while keeping assumptions visible for expert review. Its strongest role is to accelerate implementation, not to replace validation: scientific models still need to be compared with known datasets and observed outcomes. In contrast, map-based interfaces need real browser testing to ensure they work clearly and reliably for people.