Machine learningData systemsUncertaintyVerification
AI/ML Engineer building reliable intelligent systems from models to production.
Selected work
Eight worlds, in running order. Every one is drawn as its own mechanism rather than as a card, because a graph surrogate and a retrieval benchmark do not resemble each other and should not be presented as though they do. Each opens into a case study carrying its evidence and its limits.
Master's thesis / Transport
A graph network answers in milliseconds what a traffic simulator takes hours to compute. The question is not whether it is fast. It is when it may be believed.
Reliable GNN Surrogates for Transport Policy Analysis
Held-out scope100 scenarios
Enter this worldCurrent engineering / Agent security
A tool description can be scanned before anyone calls it. The document that tool returns cannot — it does not exist yet. So what is worth putting at the boundary itself, and what does it cost in refused legitimate calls?
The boundary a scanner cannot stand at
Attacks caught92.3%
Enter this worldCurrent engineering / Synthetic model
One captured source becomes an index, a graph and a cache. Days later, is any of them still telling the truth about it?
Public-safe modelSynthetic throughout
Enter this worldResearch prototype / Medical imaging
A chest report that never mentions a finding has not ruled it out. Training a classifier as though it had is the mistake this project is about.
Published performanceNone
Enter this worldReference implementation / Retrieval
Three policy forms differ by a few words. The nearest passage can be the correct clause taken from the wrong document, and read exactly like the answer.
InsureAssist: A Measured RAG Benchmark
Form discrimination0.556 top-document accuracy
Enter this worldReference implementation / Delivery
Promotion is not the next box along a diagram. It is a conjunction of four checks, and one short check stops the candidate where it stands.
A Testable End-to-End MLOps Pipeline
Held-out accuracy0.8067
Enter this worldTUM project seminar / Hydrology
Disturb the rain and a calibrated model barely notices. Disturb the ruler the river was measured with and the calibration collapses.
Uncertainty Quantification in Hydrology
Best seminar resultNSE 0.908
Enter this worldScientific computing / Benchmark
A leaderboard invites one reading: rank the rows. These three were not scored on the same task, and the gap between the questions is wider than the gap between the models.
Synthetic Streamflow Forecasting Benchmark
One-step benchmarkR² 0.979
Enter this worldDeep learning experiment / Vision
Sixty-four percent is the mean of ten very different numbers, and the boundary the model learned best is not one of the ten it was asked for.
CIFAR-10 CNN: A Reproducible Baseline
Tracked test accuracy64.26%
Enter this worldSupporting work
Smaller than the flagships and built the same way. Reading all nine READMEs side by side turned up the only fact worth leading with: 2 of them publish tracked numbers, 2 establish a working path without claiming accuracy, and 5 publish no metric at all — each one naming the artifact it would need first. Those five are drawn as the empty charts they are.
River discharge is well understood physics and expensive computation. This is the deep-learning shortcut, measured on a catchment that can be regenerated anywhere.
Thirty days of precipitation, temperature and soil moisture, produced by a deterministic rainfall-runoff generator written for the repository.
An LSTM reads the window and predicts the next day's discharge in a single forward pass, trained with early stopping and a learning-rate schedule.
The catchment is synthetic. Every score describes the generated benchmark, and none of it is real-catchment performance.
Flood events are the 95th percentile of the generated series. Detection is markedly harder than the fit, and the repository publishes both rather than the flattering one.
Four families of model on the same generated task, compared under cross-validation instead of asserted.
Five thousand samples from a seeded generator whose features are deliberately class-correlated.
Logistic regression, random forest, XGBoost and an RBF SVM run through the same pipeline, with cross-validation and hyperparameter tuning.
The labels are generated, not laboratory measurements. These scores measure how well each model recovers the generator's own distributions and say nothing about whether real water is safe to drink.
Tuning moved XGBoost from 0.896 to 0.898 and left the untuned SVM ahead. The comparison is reported as it came out.
A model behind a typed HTTP contract. What it establishes is that the path works, not that the forecast is right.
A calendar year and month, validated at the boundary before anything reaches the model.
A FastAPI service loads a regression model trained on Munich traffic-accident records and returns a rounded predicted count.
An educational prototype, not a public-safety forecast. The tracked model and CSV come from a historical-data exercise and establish no accuracy for any future decision.
A working request-to-prediction path is a real engineering result and a small one. It is reported as such, with no accuracy figure attached to it.
An exploratory analysis that is scrupulous about what it is exploring: its own synthetic dataset, and nothing else.
Ten years of daily observations for six cities, produced by a fixed-seed generator with per-city climate parameters.
Wrangling, statistical summaries, seasonality and correlation, rendered as a set of figures.
Not weather-station observations, climate evidence or forecasts. Nothing here supports an operational, scientific or policy decision.
The seasonality the analysis finds is the seasonality the generator was given. That is a check on the pipeline, not a finding about climate.
A lithium-ion cell carries no fuel gauge. Six approaches to inferring one, gathered so they can be compared instead of trusted.
Voltage, current, temperature and cycling history, from authorised NASA battery files or a deterministic synthetic generator.
SVR, random forest, XGBoost, LightGBM and an LSTM regress state of charge, alongside clustering for operating regimes and a genetic-optimised fuzzy estimator.
No dataset, trained weights or tracked evaluation are included. It must not be used to operate a battery-management system or make a safety decision.
Earlier score tables and degradation claims were removed on purpose. The repository has no versioned source split, run configuration, model artifact or metric report to substantiate them, and it names exactly what a meaningful benchmark would have to record: cell identifiers, data version, preprocessing parameters, temporal split, seed, dependency versions and evaluation artifacts.
Predicting a claim is half the problem. The probabilities also have to mean something, and someone has to explain them.
Policy and vehicle columns from a Kaggle claims dataset, fetched by the user under its own terms and never versioned here.
Encoders and scalers fit on training data only, then classifiers, Platt or isotonic calibration, cost-sensitive threshold selection and SHAP attribution.
The example cost matrix is illustrative and is not a validated business policy. Nothing here may be used to make automated decisions about people or policies.
The repository contains source and notebooks only. It versions no data, no split, no trained model, no calibration output and no evaluation report, so there is nothing to substantiate a number with.
Two tracks on one task, evaluated the same way, so the comparison is worth something when someone runs it.
AG News headlines fetched at run time. If access fails the pipeline stops rather than quietly evaluating template-generated text.
TF-IDF with logistic regression, linear SVM or random forest on one side; DistilBERT fine-tuned through the Hugging Face Trainer on the other.
The offline synthetic fixture exists for development only. Its results are not comparable to AG News or to real news classification.
Data, trained models, checkpoints, metrics, plots and run metadata are all untracked, so the repository declines to state how the two tracks compared.
Identical weights always behave identically. The interesting question runs the other way, and it is the subject of a TUM mathematics seminar.
Small fully connected networks with tanh, sigmoid or ReLU activations, constructed to be related by hidden-unit permutations and sign flips.
Numerical diagnostics look for exact and near-exact clone pairs, sampled non-degeneracy and parameter alignment, then test proposed symmetry-breaking regularisers.
The activation labels describe the assumptions this prototype considers. The formal statements live in the cited papers, not here.
The checks are finite numerical diagnostics on sampled inputs. They can expose a symmetry; they cannot prove global functional equivalence or satisfy the hypotheses of a published identifiability theorem, and the repository versions no experiment configuration or trained model.
Demand forecasting, supplier dependability and how much stock to hold, in one dashboard, with each proxy named as a proxy.
Order records from the DataCo dataset, downloaded under its own terms and deliberately ignored by version control.
Cleaning and KPIs, a comparison of demand-forecasting baselines, and the classical EOQ, safety-stock and reorder-point calculations, served through a Plotly Dash app.
The inventory values are illustrative outputs whose assumptions must be validated before they touch a real operation, and the proxy must never be read as a supplier-quality assessment.
No dataset, processed data, forecast artifact or verified business metric is versioned. The source carries no supplier identifiers and no defect measurements either, so supplier analysis groups departments as an explicit proxy and derives its quality-like score from late delivery.
Current engineering
The shape of the work I do now, drawn synthetically. It describes a class of problem common to data platforms rather than any particular system, and every figure in it is an illustrative model.
Ingest
Represent
Verify
Observe
Illustrative model. Synthetic throughout; it describes a class of problem rather than any particular system.
Research
A surrogate that is wrong quickly is worse than a simulation that is slow. The research record is about the evidence that closes that gap.
Master's thesis submitted at Technical University of Munich — Mathematics in Science and Engineering.
Experience
Each role is written as the class of technical problem it put in front of me, rather than as a job title with a date range attached.
Now
Records, vectors and a graph all describing the same material. Each is useful, each can drift, and agreement between them has to be demonstrated rather than assumed.
Research
Learning a surrogate is the easy half. Establishing when its answer may be trusted, and building the machinery that says so, is the work.
Earlier
Several spreadsheets, reconciled by hand on a schedule. The interesting part was not the automation; it was making the consolidated figure reproducible.
BP-IT Consulting & Solutions GmbH
AI Engineer (Working Student)
Built verification workflows for a multilingual legal knowledge platform using relational, vector, and graph storage.
Technical University of Munich
Student Research Assistant / Programming and Visualization
AUDI AG
Intern, Programming of Workflows and Linking of Databases
Technical University of Munich
Student Research Assistant, Numerical Methods and Scientific Visualization
IISER Bhopal
Summer Research Intern, Machine Learning for Li-ion Battery State Estimation
Build systems that can explain what they know.
And, just as importantly, systems that can say where they stop knowing it. If that is the kind of engineering you need, the case studies and the research record are the fastest way to judge whether it is any good.
Open to full-time Machine Learning and Applied AI roles