Whale ID · Film

Origin

Conservation infrastructure is invisible by default. The Whale ID model could be the most rigorous humpback re-identification on the water and still die in obscurity if its only artifacts were a GitHub repository and a journal submission. Most of the world doesn't read journals. Most funders, partners, and policy actors don't read repositories either. Some part of the work has to be legible at film tempo.

Problem

How do you communicate scientific machinery without flattening it into spectacle? Most science films either reduce a method to a slogan or hide the method behind a soundtrack. Neither approach reflects what was actually built.

Approach

Twenty-one-shot Remotion sequence built from raw Moorea fieldwork — boat, hydrophone, individual whales surfacing, the model's matching process — composed in sequence. Edited as a short film, not a demo reel. The viewer sees what was photographed, then watches the system find that animal in a corpus of others.

Methodology

Remotion lets the visual sequence be code. Frame timing, opacity curves, and audio sync are version-controlled. That matters because the film has to evolve as the model does — when the threshold changes, when the corpus grows, the cuts have to update without losing the through-line.

Selected milestones

• 21-shot sequence assembled
• In active production for late-2026 release

Open questions

• What the audience for a film about a model actually is
• How to render confidence honestly — a 91% match looks identical to a 99% match on screen

Ask me about

• How the model output is rendered without losing scientific honesty
• Why Remotion (versioned visual code) over a traditional NLE
• Who the film is actually for

Stack