Skip to main content

Spatial Intelligence Using Hexagons

· 3 min read
Sina Kashuk
Sina Kashuk
CEO
Isaac Brodsky
Isaac Brodsky
CTO

Sina Kashuk, CEO of Fused, and Isaac Brodsky, CTO and co-founder of Fused, both worked on H3 at Uber. In this talk at the H3 Conference 2026 in New York City, they revisited the H3 ideas Sina presented eight years ago and showed two new things built on them: an AI model that scores hexagons offline on a laptop, and an open H3-native climate dataset.

Eight years later, the same ideas​

Eight years ago, Sina presented H3 at Uber's office around the time it was open-sourced. His first commit to Uber's main H3 repository was an example of how to unify data with it. For this talk he gave an AI a short prompt asking it to rebuild that 2018 presentation, and it produced a new version of the same material.

The core concepts have not changed: H3 as a unified data layer, and its hierarchy of resolutions. What has changed is some of the API naming. The function Sina used in 2018 to convert a latitude and longitude to a cell, geoToH3, is now latLngToCell. He asked the AI to list the API changes since 2018, and it handled the migration in one shot.

Ring-based kernels for smoothing​

The main demo from the original talk is convolution-style analysis on H3. Because every cell is surrounded by rings of neighbors, you can assign a weight to each ring and spread a value outward, turning your data into a smooth heat map.

Sina's kernel used the center cell plus three rings around it, and a chart beside the map showed the weights as half of a bell curve, a normal distribution over distance. When he raised the weight of the outermost ring, the heat map recalculated in real time and stopped looking normal, because distant cells now carried more weight than nearby ones.

Scoring the vibe of a hexagon​

The night before the conference, Sina used AI to build a second app. It takes Overture places, looks at the category of each point of interest, groups them into hexagons, and asks what the "vibe" of each hexagon is.

Sina writes a sentence, and a new AI classification model compares it with the categories found in each hexagon to produce a similarity score. The model takes about 10 ms per classification. After loading once, it worked through every hexagon on stage one call after another, in series rather than in parallel. He asked two questions:

  • Where are the touristic areas in the city?
  • How dog friendly is each area?

The app runs in Fused Render, a new Fused product that is 100% local software. You install it on your own computer and run models like this for free, with no data center involved. Sina noted it keeps working with Wi-Fi turned off.

100 TB of H3-native climate data​

Isaac closed the talk by retrieving hourly climate data for San Francisco from two different years: several hundred thousand observations for a single location, loaded in a few seconds.

The access layer is something Fused built to make open geospatial data quick and easy to reach. The data itself is hosted on Source Cooperative. It is a new dataset of ERA5 and gridMET climate data, on the order of 100 TB, published in an H3-native format as open data.

Learn more​

Jump to section


← Back to blog