Skip to content

MagicLex/awesome-ml-systems

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

19 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

awesome-ml-systems

awesome-ml-systems

systems Hopsworks

One small, honest ML system per day, each built end to end on Hopsworks. Same shape every time: an FTI (feature, training, inference) pipeline, a real result with its caveats, and a served model you can poke at. No notebooks-that-never-ship, no accuracy without a holdout, no demo wired to a mock.

The series

# system the question result published repo
001 README Vaporware Score does a repo get abandoned, from its README text alone? ROC-AUC 0.76 2026-06-29 readme-vaporware-score
002 Asteroid Doomsday-o-meter how big is an asteroid (so, how dangerous), from its Gaia spectrum alone? size error ×1.13 vs ×1.34 blind 2026-06-30 asteroid-size-from-light
003 Phishing at Issuance is a freshly issued TLS certificate phishing, from its hostname alone? ROC-AUC 0.78 holdout vs 0.50 blind 2026-07-01 phish-at-issuance
004 Where on Earth which country was a photo taken in, from its pixels alone? top-1 52.3% / top-5 79.8% over 173 countries vs 21.2% zero-shot 2026-07-02 where-on-earth
005 How Predictable. can a machine learn your taste in 30 clicks, live, in front of you? crowd prior 0.719 pairwise vs 0.511 zero-shot; per-user Bayesian layer climbs on-screen 2026-07-03 how-predictable
006 Live Sky Watch where will every aircraft over Europe be in 60/180/300 s, and which one is not behaving like traffic here? live same-sample: model 964 m vs physics 1427 m at 60 s where it intervenes; jamming grid + learned normalcy 2026-07-06 live-sky-watch
007 Ghost Fleet which vessels behave like the sanctioned shadow fleet, from their AIS tracks alone? 9.4x lift over a blind sanctions-list lookup, ROC-AUC 0.92 (population split); live network reveal 2026-07-07 ghost-fleet
008 the untested which never-tested plant might fight a drug-resistant infection, from molecular structure alone? mean AMR ROC-AUC 0.80, beats 1-NN Tanimoto on every scored head; recovers Artemisia for malaria from structure alone 2026-07-08 the-untested
009 downwind what is in the air where nobody is measuring? PM2.5 20.9% RMSE under the raw CAMS prior at leave-stations-out stations (r2 0.61 vs 0.38); live all-Europe field with a monitored-vs-predicted frontier 2026-07-09 downwind

The dog house

One honest exception. The series test is a decision someone can act on. These builds are clean FTI systems and nice to look at, but the use case does not hold up to that test, so they sit here, unnumbered, kept public because the engineering is real.

system the question why it is here repo
dead-air can you hear a solar flare black out the shortwave bands before the bulletin? beautiful instrument, no decision attached: a few-minute lead measured against a bulletin nobody waits on (and negative on the holdout), and it misses roughly four flares in five. You cannot act on the ionosphere anyway. Real code, wrong problem. dead-air

The standard

Every repo in the series follows the same mould, so they read as siblings.

Shape. An FTI system on Hopsworks. Sources to a feature pipeline to a Feature Group, a Feature View to training to the Model Registry, a deployment to an endpoint, an app that calls it. The skeleton lives in templates/diagram.mmd.

Banner. Generated, not hand-drawn, so 30 of them stay consistent. Dark canvas, emerald accent, the Hopsworks hop-mark as the fixed brand, only title/tagline/emoji/index change per repo.

python tools/make_banner.py \
  --title "My System" \
  --tagline "What it predicts, in one honest sentence." \
  --emoji "🧪" --index 002 --out assets/banner.svg

README. Result first (with the metric and the holdout), then caveats, then architecture (the diagram plus a file-by-file map), then reproduce, then the served demo. Start from templates/README.template.md.

Honesty rules. The label is named and its proxy is stated. There is a holdout number, not just cross-validation. No feature leaks the label. Heavy fits run as Hopsworks jobs, not in a terminal. Feature extraction is one shared function so training and serving cannot skew.

New entry

mkdir ../my-new-system && cd ../my-new-system
cp -r ../awesome-ml-systems/tools .                       # the banner generator
cp ../awesome-ml-systems/templates/README.template.md README.md
python tools/make_banner.py --title "..." --tagline "..." --index NNN
# fill the README, paste templates/diagram.mmd, then add a row to the table above

About

One small, honest ML system per day, built end to end on Hopsworks.

Topics

Resources

Stars

3 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors