Hi HN, Alex here, designer-engineer in the Netherlands. My day job is
Repper (a pattern-design tool). Skymap is a side project, built last
weekend in 4 days (~32 hours, AI-assisted with Claude Code).
I've followed Anton Petrov on YouTube for years and kept seeing nice
pans over galaxy catalogs in research videos, thinking "why can't I just
fly through that?" Aparently nobody had built it for the browser yet.
It loads SDSS (~500k galaxies), 2MRS (~45k), and GLADE (~2M) as WebGPU
instanced billboards. Past a certain on-screen size each galaxy
crossfades from a dot into a procedural oriented disk (catalog
axis-ratio + PA from HyperLEDA / 2MASS XSC), then into a real survey
image when close enough: SDSS DR18 ImgCutout, or DSS via CDS hips2fits
for the rest. There's a cosmic-web filament overlay built offline with
DisPerSE (Sousbie 2011), a faint blue lattice over the point field.
Striking at supercluster scale.
What I genuinely like about coding with Claude is that I can read about an
algorithm in a paper and have it running interactively the same
evening. DisPerSE skeletons, Schechter LF corrections, HEALPix
re-weighting, K-correction in the fragment shader. Stuff I'd otherwise
only have watched a video about.
The interesting texture across these 4 days wasn't bug-fighting, it
was iteration: continuous refactoring, guiding Claude away
from slop, accepting that scaffolding gets re-done. The renderer
wasn't new ground (years of WebGL behind me). Deep-link
resolution and tier swaps end up as a useEffect-and-refs web I'd
normally write as redux-saga; I went fully agentic this round and
you can feel the absence of an orchestrator (ouch).
UX has plenty of rough edges, mobile gestures aren't there yet. I'm not
an astronomer (the cosmology math is all Claude-derived against
textbook formulas), so if anyone here works in the field and spots
something off, I'd love to hear it.
I'm also curious how people grow a side project from a code seed to something
that is robust enough to be used in production, especially with agentic
coding. I feel I learned a lot about this, but I'm curious about your
experience. TBH, I like my code clean, so with the speed this is going
I end up with a lot of cringe code that I'd wish would be better.
really good
Very cool! The magnitude of how tiny we actually are...
Thanks! Curious if it works well on your machine; I did some work optimizing. Are you using desktop or mobile?
Works well for me. I'm on desktop (Macbook Air M3) but have a decent amount of RAM
Good to know. Have you tried switching to the large data set?
Right! While building this I found out this dataset only spans 2% or so of the actual observable universe...
Hi HN, Alex here, designer-engineer in the Netherlands. My day job is Repper (a pattern-design tool). Skymap is a side project, built last weekend in 4 days (~32 hours, AI-assisted with Claude Code).
I've followed Anton Petrov on YouTube for years and kept seeing nice pans over galaxy catalogs in research videos, thinking "why can't I just fly through that?" Aparently nobody had built it for the browser yet.
It loads SDSS (~500k galaxies), 2MRS (~45k), and GLADE (~2M) as WebGPU instanced billboards. Past a certain on-screen size each galaxy crossfades from a dot into a procedural oriented disk (catalog axis-ratio + PA from HyperLEDA / 2MASS XSC), then into a real survey image when close enough: SDSS DR18 ImgCutout, or DSS via CDS hips2fits for the rest. There's a cosmic-web filament overlay built offline with DisPerSE (Sousbie 2011), a faint blue lattice over the point field. Striking at supercluster scale.
What I genuinely like about coding with Claude is that I can read about an algorithm in a paper and have it running interactively the same evening. DisPerSE skeletons, Schechter LF corrections, HEALPix re-weighting, K-correction in the fragment shader. Stuff I'd otherwise only have watched a video about.
The interesting texture across these 4 days wasn't bug-fighting, it was iteration: continuous refactoring, guiding Claude away from slop, accepting that scaffolding gets re-done. The renderer wasn't new ground (years of WebGL behind me). Deep-link resolution and tier swaps end up as a useEffect-and-refs web I'd normally write as redux-saga; I went fully agentic this round and you can feel the absence of an orchestrator (ouch).
Live: https://skymap.rulkens.com Repo: https://github.com/rulkens/skymap (MIT) DOI: https://doi.org/10.5281/zenodo.20037028
UX has plenty of rough edges, mobile gestures aren't there yet. I'm not an astronomer (the cosmology math is all Claude-derived against textbook formulas), so if anyone here works in the field and spots something off, I'd love to hear it.
I'm also curious how people grow a side project from a code seed to something that is robust enough to be used in production, especially with agentic coding. I feel I learned a lot about this, but I'm curious about your experience. TBH, I like my code clean, so with the speed this is going I end up with a lot of cringe code that I'd wish would be better.