0
0
0
0
0
0
0
0
0
0
0
0

· 8 min · AI/ML, Data Analytics, Career

From Android Dev to ML Pipelines: What My M.Sc. Taught Me the Hard Way

Moving from mobile engineering to AI & Data Analytics at HNU exposed every gap in my mental model. The biggest surprise: the ML part is the easy 20%.

I enrolled at Hochschule Neu-Ulm expecting to learn models. I actually learned data engineering, statistics humility, and why production ML is mostly plumbing.

The 80/20 nobody advertises

Cleaning, joining, and validating data consumed 80% of every project. The model — often a gradient boosted tree that outperformed the fancy architecture — took an afternoon.

Pandas is not a database

My first pipeline loaded 4GB into a DataFrame and died.

  • Chunked processing instead of loading the full dataset at once
  • Parquet over CSV for anything that gets re-read
  • Push filters down to the source query, not after loading

Evaluation is a design problem

Accuracy on an imbalanced dataset is a lie. Choosing the metric — precision vs recall vs calibration — turned out to be a product decision disguised as a technical one.

The best thing I built all semester was an evaluation notebook the whole team could actually read.

Why mobile devs make decent ML engineers

We already think in pipelines (data → transform → render), we already profile before optimizing, and we already distrust the happy path. The syntax is new; the discipline transfers.

← All entries · Portfolio