Victor Alejandro Alvarez Vega

This is a prototype of NSHARP that cross compiles to MacOS, Windows, Linux, and... the web browser!

It just decoded and displayed NWS and IMET upper air BUFR sounding files at full, 1 HZ resolution (4-6k vertical levels) interactively, in real time.

No runtime loading of BUFR tables needed!

A new interactive REFS (RRFS Ensemble Forecast System) viewer is coming to an SPC site near you. Probably sometime in early 2026.

New demonstration of modifying the selecting the storm motion of a sounding, including setting your own custom choices. Changing the motion changes all wind-related parameters and boxes, allowing you to see how an environment may look with a deviant storm (from twitter)

500mb windspeeds with geopotential height and wind barbs. Image from aguacerowx.com

18z HRRR upper air sounding taken from north central Illinois on March 19, 2025 at 21 UTC.

Pretty solid cold-core setup in northern and eastern Illinois tomorrow. Impressive moisture return will allow for low to mid 50s F dewpoints in northern/eastern Illinois. On top of this moisture plume, ~-20C temperatures will allow for strong low and mid level lapse rates to develop.

I have completely remade sharppy in javascript. This is for an upcoming model viewing program and im super happy with the current result of it. sounding from 5/31/2013, at Powers OK. (from twitter)

AguaceroWx FREE OPEN BETA is LIVE for limited time! Already have an account? Go to aguacerowx.com/app. Need to register? Visit aguacerowx.com/register. We welcome your feedback!

Join our discord here to get updates/make suggestions! discord.gg/HxkzgbrFjt

I’ve been rewriting NSHARP, our sounding analysis visualization program. It’s fully interactive at 140+ FPS, supports full-res/raw 1s radiosonde data natively. It compiles in MacOS, Linux, Windows, and runs in a web browser. Of course, leveraging the sounding code above!

4/?

I'm excited to put this into experimental, and eventually operational, use over the coming weeks as we switch back over to client-side development before going fully online again. Follow along at @aguacerowx.bsky.social for future updates as we get closer to being back online!

This offers several benefits: unused tiles are never generated, tiles can now be created dynamically, the tile format is optimized for web use, and near-unlimited instant horizontal scalability matches well with the bursty nature of client visualization needs.

There is a tradeoff here: the first user to request a tile will experience this slight delay. However, one of the primary benefits of MVT is its ease of caching, meaning future users will simply request the pregenerated tile. This is what this looks like below.

When a user requests a field, each tile on the map triggers a unique Lambda. This requests the bounding indices in the stored array. Contouring and encoding to MVT from this point forward is light and fast, the full process on the most complex geometries completing within 1-2 s.

The solution was to embrace serverless. Instead of processing GRIB -> geometry -> complete tile pyramid in one go, we compute desired fields that are non-native to GRIB and keep the data in array format. Any further processing from this point forward is event-driven.

In short, the unique domain challenge with vector tiling and model data visualization is scale. Model data contains many different fields across forecast hours, runs, and models, quickly becoming expensive to pre-process entirely.

Designing a Serverless Backend for Data Processing with Event-Driven Vector Tiling A little over two months ago today, data processing operations on Aguacero were stopped as we initiated our migration to Amazon Web…

I took on this challenge and after endless hours of research and experimentation, I want to share the technical journey of our solution in the below Medium article.

medium.com/@victoraalva...

Since late September, me and @MikeBWeather (on Twitter) have been hard at work moving Aguacero over to AWS. During this process, it became clear that we would need to reengineer our data processing pipeline.