I'm Alex, an artificial intelligence student at THD in Deggendorf, Germany. These are my AI and software projects, with the work, results and links behind each one.
My portfolio is also a game: sail Harold the rubber duck through the flooded internet of 1999. You can read the same projects here without starting the game.
CheXVision
two models looked at 112,120 chest x-rays. the borrowed one did better.
a four-person university project on the nih chestx-ray14 set: fourteen pathologies, one picture each.
we trained a residual network from nothing and a densenet-121 borrowed from imagenet, under the same protocol. the borrowed one reached 0.846, a hair above the published chexnet 0.841. the one we built ourselves reached 0.814.
educational research. not a clinical product, and it says so on every page.
x-rays
112,120
macro auc-roc
0.846
from scratch
0.814
Titanic ML Platform
five docker containers, one iceberg. a team of seven, and it still floats.
a university team of seven, under the name random-iceberg. the brief was the titanic dataset; the work was everything around it.
five services: frontend, web api, model api, postgres, auth. sign in, train a model, ask it who survives, keep your history.
i set the architecture, wrote the proposal and the user stories, and reviewed every pull request.
services
5
crew
7
prs reviewed
all
THD Room Finder
which study rooms are free right now, and for how long. no login.
thd publishes its timetable. it does not tell you which room is empty, or until when.
room finder reads the public schedule, ranks free rooms by how long they stay free, and shows them on a small campus map. one react codebase, installable on a phone or as an android app.
the first version was two native apps. the second is one. that was the lesson.
refresh
5 min
logins
0
platforms
2
Deutsche Bahn Delay Predictor
two million train records. the board still says +12 min.
a three-person university study. the idea was the easiest one in the room: predict how late a deutsche bahn train will be.
the work was not easy. we cleaned 2,061,357 records, chose 16 features by forward selection, and compared linear regression, k-nearest neighbours and a random forest on held-out data. the random forest won, its test error 92.4% below a naive baseline.
the paper is written in quarto and rebuilds itself from the notebook. the train on this island has not arrived yet.
records
2.06m
features
16
vs baseline mse
-92.4%
Waste Sorting Assistant
four kinds of german bin. a camera that knows which is which.
international students in deggendorf get a bin system explained only in german. biomüll, glas, papier, restmüll, and fines when you guess wrong.
three of us walked around town and photographed 466 real bins, labelled them by hand, and fine-tuned yolov8 on them. it runs live on a phone camera through a streamlit app.
i did the devops and the deployment. the lesson: start from a pretrained model, but photograph your own bins.