I build machine learning systems and the services that run them. My Master's thesis at the Technical University of Munich added calibrated confidence estimates to a graph neural network that predicts how Paris traffic responds to road policy. At BP-IT Consulting I built retrieval services over vector and graph databases, a document extraction pipeline that pairs OCR with large language models, and the ML core of a chest X-ray classifier. I work the whole path from data and training to a tested, deployed service, and I measure what a model gets wrong as carefully as what it gets right.
View the full diagram, which adds a panel of the real data at each stage and names the file every number is checked against. Dark version.
The ring is the shape of it: nine stages, three bands that hold through all of them, three ways back. The diagram under it is the same nine stages laid out as the system they are, with the symbols each one works with, the order the work runs in, and the three arrows that go back. Each of those is labelled with the reason it fires, because a loop that never says why it returned is a circle.
Three bands hold underneath all nine: documentation and reproducibility, privacy and governance, ethics and bias. The track across the top shows how far a data analyst, a data scientist and an ML engineer usually go. An analyst and a scientist reach the same stage and differ in depth rather than in reach, which is why two of those three bars are the same length.
The full diagram adds what this one leaves out: under each stage, a panel of what the data actually looks like at that point, from raw rows as the city exported them through a typed table, a schema, a five-element feature vector, a prediction with its measured coverage, a promotion gate and a drift reading. Every one of those panels is real and every callout names the file it is checked against. It is drawn at 1860 pixels wide, which is more than a README column can show, so it is linked rather than inlined.
Both versions above animate. The full one also has a still copy in the repository, light and dark, for anywhere that does not play animation.
Legal knowledge base, BP-IT
German and EU legislation parsed into PostgreSQL, embedded into Qdrant and linked in Neo4j, with a verification layer that decides what counts as certified.
52 laws certified complete: 32 German and 20 EU
I built four of the five services. The Ingestion Preprocessor, which parses and normalises the law XML and owns the canonical PostgreSQL. The Ingestion Loader, the only writer into Qdrant and Neo4j. The Legal KB Health Service, which runs up to 14 gates per law. The Legal KB Dashboard, which shows what passed and what did not. I extended the fifth, the Entity Producer, which predates legal ingestion.
Chest X-ray classifier, BP-IT
A DenseNet-121 over the 14 NIH findings, learning from three datasets that each annotate a different subset of them.
22 of 42 source and finding pairs carry a label. The rest are masked out of the loss rather than counted as absent
I built the training script, the split module, the leakage audit, the tests, the benchmark and the AI Radiologist demo. Not mine: the three datasets, which belong to Stanford, the NIH Clinical Center and Kermany and colleagues, and the pre-trained backbone the demo runs.
Release tracking at AUDI AG
An Excel VBA application with UserForms replacing a manual search through a development release sheet, wired to the release system every department already used.
One click in place of a scan through roughly 600 rows and 130 or more columns, with separate German and Chinese market interfaces
I built the whole application: the UserForms, the lookup over the release sheet, the write-back, and the separate German and Chinese market interfaces. Not mine: FREIPRO, the release system it connects to, which every department already used.
Nothing from the confidential draft appears here or in that repository. The AUDI case study carries no AUDI code, no internal data and no screenshots.
Uncertainty for a traffic surrogate
A graph network predicts how Paris traffic shifts when a lane closes, in milliseconds instead of hours. The year went into whether it knows when it is wrong.
82% less calibration error, and 41.2% less error on the half it keeps
I built the uncertainty layer: the calibration, the conformal intervals and the selective-prediction evaluation. Upstream work by Elena Natterer and colleagues: the PointNetTransfGAT surrogate, the MATSim preprocessing and the base training pipeline it sits on.
Uncertainty in a hydrological model
A TUM seminar on what a calibrated model loses to the uncertainty around it, in the parameters and in the measurements.
0.148 NSE lost to observation error, far more than rainfall noise costs, and recalibration recovers only 5.76% of it
A three-person group submission, and the science is the group's. My part was the HBV and Cython integration, the data configuration, the production runs behind the published results, and the report assembly.
Identification of neural networks
A TUM research seminar on Fefferman's framework: when do the weights of a network follow from the function it computes, and when do they not.
A network and its permuted, sign-flipped twin agree to 1.19e-7. Same function, different weights. tanh is identifiable, ReLU is not
I built the constructions, the permuted and sign-flipped twins, and the numerical diagnostics that expose them. Not mine: the framework, which is Fefferman's, and the seminar was supervised at TUM.
A gateway that refuses
Nine controls at the three points where an MCP tool call is untrusted, and a deterministic benchmark that measures which control stops which attack.
96.2% of attacks caught at 9.5% false positives, against a keyword filter's 38.5%. Three scored misses stay in the corpus
I built all of it, the nine controls and the benchmark that measures which control stops which attack.
Retrieval that cites its sources
Hybrid retrieval over NFIP flood-insurance policy forms, with a benchmark that isolates which change earned which gain.
Top-document accuracy 0.556, 3.3x from the 0.167 starting point, earned by the chunk size. Fusion earned the ranking, MRR 0.365 to 0.420
I built all of it, the hybrid retrieval and the benchmark that isolates which change earned which gain.
Drift in the target, not the features
Hourly demand forecasting that watches its own target for drift and retrains itself. A feature-only monitor would have missed the thing that moved.
0.679 target drift while only 9% of features move, and a feature-only monitor sleeps through a 63% rise in demand
I built all of it, the target-drift monitor, the retraining trigger and the service around them.
Forecasts that argue with themselves
Monthly Munich road-accident forecasts behind a service. Re-run on an ordinary pre-pandemic period, the published margin does not survive, and the repository says so.
A published +23.4% margin becomes -15.5% on ordinary years. CI reproduces the published run to 1e-9 or fails
I built all of it, the forecasting service and the re-run that contradicts the published margin.
A promotion gate that can refuse
The whole lifecycle for one small task: validation that can stop the run, leak-free feature fitting, immutable bundles, and a gate that requires a margin over a baseline.
0.8795 ROC-AUC on the held-out split, behind a gate that can say no
I built all of it, every stage from validation to the promotion gate that can refuse.
| Languages |
|
| ML and DL |
|
| GenAI and Retrieval |
|
| MLOps and Cloud |
|
| Backend and Data |
|
|
|
|
Two badges are shortened to fit the strip. The full titles are Improving Deep Neural Networks: Hyperparameter Tuning, Regularization and Optimization and Data Science: Python for Data Analysis Full Bootcamp, which is what each verification page shows.
16/16 figures on this page were re-checked against the file that produced them, by the workflow that built this line.
