WildVector: Making Migration Data Legible to Students

Inspiration

I was inspired by how LLMs can put large, messy scientific datasets directly into students' hands. Movebank alone holds millions of GPS fixes from tagged animals, but a CSV of lat/lon pairs means nothing to a classroom. I wanted to create a visualization tool, paired with a robust mathematical underpinnings that could help make sense of this data and inspire kids to learn more.

What it does

WildVector is a Streamlit app where students explore real Movebank migration telemetry for turkey vultures, Arctic foxes, and blue whales, then run "what-if" weather experiments and watch a live-animated map compare the recorded baseline journey against a model-projected scenario route, colored consistently so it's always clear which line is "what happened" versus "what the model predicts."

How I built it

I built this iteratively using GPT 5.6: a blueprint and implementation plan up front, then repeated cycles of prototyping capped off with a dedicated pass for performance and polish.

Under the Streamlit/pydeck UI:

• Corridor construction. Each animal's journey isbroken down into progress stations. At each station, the centerline is a weighted median latitude and a weighted circular mean longitude.
• Reweighting, not fabrication. A scenario never invents coordinates — it scores each recorded journey against a validated effect and reweights it by exponential tilting, then recomputes the same weighted-median corridor over the same animals.
• A statistical activation gate. Each outcome model is trained with rolling-origin cross-validation and only allowed to influence what students see if it beats seasonal-median and persistence baselines.

Challenges I ran into

The two hardest problems weren't rendering a map — they were trustworthiness and speed.

Trustworthiness: it's trivially easy to build a slider that produces a smooth, convincing, fabricated route. I gated this at both build time and runtime — if a shipped experiment doesn't have a model that passed validation, the app itself refuses to render rather than show an unvalidated effect. That gate exposed a subtler bug: the outcome model still reports its supported effects even for a typical (zero-delta) scenario, so a baseline preset could carry scenario-colored markers it shouldn't. The fix was gating display purely on whether the selected scenario is actually non-typical, not on whether the model "has an opinion."

Speed: live playback was recomputing the entire routing pipeline on every animation tick. I split the expensive, per-session-constant geometry from the cheap per-frame marker interpolation, cached the former by content across reruns, and pre-warmed every species/experiment combination at launch.

Accomplishments I'm proud of

The method I'm most proud of is the reciprocal season transfer. Telemetry is asymmetric — a species might have enough fall journeys to validate a model but too few spring ones. Instead of showing nothing, or silently reusing the fall model as if independently proven, _transfer_route_deformation borrows the shape of the validated season's response and projects it onto the other season's known route, tagged support = "paired-season projection" so the UI always distinguishes measured from inferred. It's a small piece of statistical honesty that matters in a tool meant to teach kids what evidence looks like.

What I learned

A data tool for education fails differently than a normal dashboard: a beautiful wrong answer is worse than an ugly missing one. I learned to design the modeling layer around refusal and that in an interactive geospatial tool, correctness under real-time constraints is its own discipline: a routing pipeline that's fine once per page load is a real architecture problem at 5 frames per second.

What's next for WildVector

More datasets and modeling and getting it in front of local teachers for feedback and integration into lesson plans. The app as designed can be made to work for anything kindergarten to college. All that changes is how, and what parts of the app you interact with.

Built With

• duckdb
• pydeck
• python
• scikit-learn
• streamlit