LGND wants to make ChatGPT for the Earth

Image Credits:NASA

La Tierra está inundada de datos sobre sí misma. Cada día, los satélites capturan alrededor de 100 terabytes de imágenes.

Pero hacer sentido de ellos no siempre es fácil. Preguntas aparentemente simples pueden ser complicadas de responder. Tomemos esta pregunta que tiene una importancia económica vital para California: ¿Cuántas barreras contra incendios tiene el estado que podrían detener un incendio en su camino, y cómo han cambiado desde la última temporada de incendios?

“Originalmente, tendrías a una persona mirando fotos. Y eso solo se puede escalar hasta cierto punto,” Nathaniel Manning, cofundador y CEO de LGND, le dijo a TechCrunch. En los últimos años, las redes neuronales han hecho que sea un poco más fácil, permitiendo a expertos en aprendizaje automático y científicos de datos entrenar algoritmos para ver barreras contra incendios en imágenes satelitales.

“Probablemente inviertas, sabes, unos pocos cientos de miles de dólares — si no millones — para intentar crear ese conjunto de datos, y solo podría hacer eso una cosa,” dijo.

LGND quiere reducir esas cifras en una orden de magnitud o más.

“No estamos buscando reemplazar a las personas que hacen estas cosas,” dijo Bruno Sánchez-Andrade Núñez, cofundador y jefe científico de LGND. “Estamos buscando hacerlas 10 veces más eficientes, 100 veces más eficientes.”

LGND recently raised a $9 million seed round led by Javelin Venture Partners, the company exclusively told TechCrunch. AENU, Clocktower Ventures, Coalition Operators, MCJ, Overture, Ridgeline, and Space Capital participated. A number of angel investors also joined, including Keyhole founder John Hanke, Ramp co-founder Karim Atiyeh, and Salesforce executive Suzanne DiBianca.

The startup’s core product is vector embeddings of geographic data. Today, most geographic information exists in either pixels or traditional vectors (points, lines, areas). They’re flexible and easy to distribute and read, but interpreting that information requires either deep understanding of the space, some nontrivial amount of computing, or both.

Geographic embeddings summarize spatial data in a way that makes it easier to find relationships between different points on Earth.

“Embeddings get you 90% of all the undifferentiated compute up front,” Núñez said. “Embeddings are the universal, super-short summaries that embody 90% of the computation you have to do anyways.”

Take the example of fire breaks. They might take the form of roads, rivers, or lakes. Each of them will appear differently on a map, but they all share certain characteristics. For one, pixels that make up an image of a fire break won’t have any vegetation. Also, a fire break will have to be a certain minimum width, which often depends on how tall the vegetation is around it. Embeddings make it much easier to find lugares en un mapa que coincidan con esas descripciones.

LGND has built an enterprise app to help large companies answer questions involving spatial data along with an API which users with more specific needs can hit directly.

Manning sees LGND’s embeddings encouraging companies to query geospatial data in entirely new ways.

Imagine an AI travel agent, he said. Users might ask it to find a short-term rental with three rooms that’s close to good snorkeling. “But also, I want to be on a white sand beach. I want to know that there’s very little seaweed in February, when we’re going to go, and maybe most importantly, at this time of booking, there’s no construction happening within one kilometer of our house,” he said.

Building traditional geospatial models to answer those questions would be time consuming for just one query, let alone all of them together.

If LGND can succeed in delivering such a tool to the masses, or even just to people who use geospatial data for their jobs, it has the potential to take a bite out of a market valued near $400 billion.

“We’re trying to be the Standard Oil for this data,” Manning said.