Sample work · AI & Automation
Machine Learning: Tabular model with honest comparison and ready delivery
A small neural network on sample data, measured against random guessing, logistic regression, and gradient boosting, with cross-validation. Plus delivery: model in three formats and a prediction script that produces identical results in a fresh process.
Sample work (demo), not a client job. Built as a trial task with invented example data; the check figures below come from the automatic test run and are not polished.
What was built
- Training with clean separation of training and test data, metrics against simple baseline methods
- Two runs with identical metrics, cross-validation shows the variance
- Delivery as state_dict, TorchScript, and ONNX with prediction script
- Countercheck: data leakage and swapped labels are detected by the checks
Download kennzahlen.json Download vorhersage.py Download ml-modell-lieferung.zip
The data is self-generated, alongside the freely available digits dataset from scikit-learn; no personal data is included. The metrics apply to this data. For customer data, I only state an accuracy after an evaluation with a separate split. CPU only, small models; large language or image models, TensorFlow, XGBoost, and time series are not tested. No operation and no hosting.
How it was checked
Automatic test run on 2026-11-01, taken from the log (test data invented, figures measured):
- Table MLP acc 0.836 F1 0.8367 (Random 0.5, LogReg 0.692, GradBoost 0.844), CV 0.8542+-0.0184, MAE 4.201 vs mean 22.021
- two runs identical metrics
- Digits-CNN acc 0.9844 (most frequent class 0.1022, LogReg 0.9622)
- Delivery: state_dict+TorchScript+ONNX
Limits
What this sample does not cover, so nobody expects more:
- CPU only (PyTorch CPU build, no CUDA/GPU); training of small models (here MLP with ~5000 weights, CNN on 8x8 images)
- small datasets (thousands of rows, images up to approx. 28x28); large image datasets (ImageNet), video, audio models not tested
- no training or fine-tuning of large models (LLM, diffusion, large transformers), no training on GPU cloud
- customer data only named with customer approval and on behalf; the sample uses self-generated data and sklearn-digits, no personal data
- metrics apply to the sample data; on customer data, an accuracy is only stated after testing with a separate train/test split, never promised in advance
- delivery as model file (state_dict, TorchScript, ONNX) + prediction script; no hosting/serving in production, no MLOps pipelines
- TensorFlow/Keras, XGBoost/LightGBM, specialized time-series models, reinforcement learning not tested
Something like this for your business?
Describe briefly what you need. You get a fixed price before I start; nothing happens without your yes.