Selected work
Research, coursework, prototypes, reference implementations, and synthetic demonstrations are labelled separately. Every case study includes evidence and limitations, not only a tool list, and every repository states what it does not establish.
01Featured case studies
Each case study states the problem, my contribution, the versioned evidence, the quality controls, and the limitations that bound the claim.
An evidence-led study of when a graph neural network surrogate is accurate, when it is uncertain, and how that uncertainty can support review decisions.
Researcher and thesis author
A retrieval-augmented question-answering service over real federal flood-insurance policy text, built so its retrieval quality can be measured rather than demonstrated.
Project author and engineer
A runnable reference for the lifecycle around a text classifier, rebuilt on a licensed dataset so the pipeline's own quality gate has something real to refuse.
Project author and engineer
A team seminar connecting HBV rainfall-runoff calibration, local and global sensitivity analysis, and input/output uncertainty propagation.
Group contributor; individual ownership of assignment results is not claimed
A compact PyTorch image-classification experiment with a tracked configuration, learning history, class-level diagnostics, and an honest reference result.
Project author
A deterministic benchmark comparing seasonal-naive, SARIMAX, and gradient-boosted one-step streamflow predictions.
Project author
An MCP client trusts three things at three different times: the tool declarations a server publishes, the arguments a model builds, and the content a tool returns. The security tooling that exists reads declarations before use, which covers the first and cannot reach the third — a poisoned search result does not exist until the search runs. This puts an enforcement point at the protocol boundary, where all three are visible, and then measures what each control is worth against a corpus that includes the legitimate traffic a careless rule would break.
Sole author of the gateway, controls, corpus, benchmark and trace format
02Engineering systems
Most data platforms end up holding the same information more than once — as records, as vectors, as a graph. Each copy is useful, and each one can drift. This is an illustrative model of how I think about keeping them honest: what gets captured, what gets derived, what happens when the source changes, and what a check actually proves.
Illustrative system model. Synthetic data throughout; it describes no particular system.
Captured before it is parsed
The bytes are retained first, so everything built afterwards has one fixed thing to be compared against — rather than being compared against another copy of itself.
Capture retained
Structured data
Ordered records with stable identities, and the relationships found in their fields.
Vector space
One embedding per record, placed so that distance stands for similarity.
Knowledge graph
A node per record and typed edges for the relationships between them.
Illustrative model, not a measurement. One writer per derived view makes a divergence attributable to a single path; it does not on its own establish that the views agree.
Rows written by the ingestion path, each with its own identity.
One embedding per record, written by the projection path.
Totals agree
Totals still agree
Both sides report the same number of records, and a dashboard turns green.
The source changes. The total does not.
Nothing about which records are present, or what they say, has been checked.
Illustrative quantities, spelled as words so they are not mistaken for a measurement. A total survives one removal paired with one addition, so equal totals establish quantity alone.
What each earlier version was built from, kept exactly as captured.
What was being served until the updated source arrived.
What is served now, after converging on the update.
The record is identical to what is held, so it keeps its identity and everything derived from it.
A record the previous version did not contain is stored and projected into both derived views.
The record keeps its place and its content is replaced, along with its embedding and its edges.
The record leaves the current version. What is removed is its place in what is served, not the record that it existed.
Current state converges on the update; what earlier versions were built from is kept apart from it rather than brought forward as though it were current. The model does not claim the update itself was correct.
Quantity
Identity
Content
Relationships
Provenance
Each comparison is recorded as its own result rather than folded into one status, because passing one is not evidence for another — and none of them turns into proof.
03How the work connects
Select any node to see what it means here and which public artifact backs it. Dashed nodes are directions of study with no public project yet.
Data
Validated, fingerprinted, versioned inputs
Models
Trained approximations of an expensive process
Retrieval & agents
Grounding a response in retrieved evidence
Production systems
Gates, registries, containers, and served contracts
Reliable decisions
Calibrated confidence and an explicit review path
Evidenced
Uncertainty that ranks likely error, calibration that survives its own protocol, conformal intervals with a stated coverage, and a review queue for the rest.
See the evidence04Repository index
The flagship repositories are shown as assemblies: one part for each focus area the registry records, plus its portfolio status and its evidence boundary. Categories describe status, not technical quality, and experiments are never presented as production systems.
Every part below is also drawn as an assembly you can scroll through and open.
mcp-policy-gateway
A proxy that sits between an MCP client and an MCP server, inspecting tool declarations, call arguments and returned content, with a 44-case corpus that measures each control against the legitimate traffic a careless rule would break.
Python
Case studymedico
A research training script that fine-tunes a DenseNet-121 across fourteen chest-radiograph findings, combining three source datasets and masking the loss wherever a label is uncertain or simply absent.
Python
Case studyml_surrogates_for_agent_based_transport_models
Master's thesis codebase studying when a graph neural network surrogate for transport simulation can be trusted, and how uncertainty supports a review decision.
Python
Case studyMLOps-End-to-End-Pipeline
The lifecycle around a text classifier on a licensed dataset: checksum-verified data that validation can refuse, leak-free feature fitting, a promotion gate measured against a baseline, atomic checksummed bundles, and a served container.
Python
Case study05Every public repository
Deep work with a written case study and inspectable evidence.
mcp-policy-gateway
A proxy that sits between an MCP client and an MCP server, inspecting tool declarations, call arguments and returned content, with a 44-case corpus that measures each control against the legitimate traffic a careless rule would break.
A prototype with a reproducible benchmark, not a deployed product. Detection is pattern-based and deterministic, so paraphrased attacks and encoded payloads pass; two of the corpus cases are kept as scored misses and two known false positives are kept for the same reason. The Model Context Protocol SDK is a dependency by other authors, not part of this work.
medico
A research training script that fine-tunes a DenseNet-121 across fourteen chest-radiograph findings, combining three source datasets and masking the loss wherever a label is uncertain or simply absent.
Experimental research code and nothing else. The repository ships no trained weights, no patient data, no held-out metrics and no clinical validation, and must not be used for diagnosis, triage or treatment.
ml_surrogates_for_agent_based_transport_models
Master's thesis codebase studying when a graph neural network surrogate for transport simulation can be trusted, and how uncertainty supports a review decision.
A fork of Elena Natterer's repository, extending it with the uncertainty work. It carries the submitted thesis as a frozen record alongside a working copy that has since been edited, and says which is which. The artifacts it reads are published separately in ml-surrogates-thesis-data.
MLOps-End-to-End-Pipeline
The lifecycle around a text classifier on a licensed dataset: checksum-verified data that validation can refuse, leak-free feature fitting, a promotion gate measured against a baseline, atomic checksummed bundles, and a served container.
A reference implementation, not a deployed product. The published result is an ordinary TF-IDF baseline on 600 held-out rows, pooled across three sources, and has never carried production traffic.
Repositories receiving current engineering attention.
insureassist-rag-mlops
A local-first insurance-policy question-answering service that retrieves source clauses and returns cited answers behind a typed FastAPI contract.
An engineering prototype. Kubernetes manifests are authored but no completed cloud deployment and no regulated-data validation are claimed.
ZQ
Source of this website: a server-first Next.js platform with a typed factual registry, evidence and privacy validation, generated metadata, and automated accessibility regression tests.
A personal platform rather than a general-purpose template. Content validation rules encode decisions specific to this portfolio.
Legitimate smaller implementations with a working end-to-end path.
DPS
A small FastAPI service that loads a regression model trained on Munich traffic-accident records and answers a calendar year and month with 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 forecast accuracy for any future decision.
DPS_Model.ipynbExplores the Munich accident recordsRegression_model.ipynbFits the regression and writes Regressionmodel.pklapp.pyValidates a year and month, loads the model relative to itself, returns a rounded countEstablishes a working path. Claims no accuracy.
Weather-Data-Analytics-EDA
A wrangling-and-visualisation walkthrough over deterministic generated daily observations for six cities, covering statistical summaries, seasonality and correlation.
Every observation is produced by a seeded generator. The figures are not weather-station records, climate evidence or forecasts, and support no operational or scientific claim.
src/generate_data.pyWrites ten years of daily observations for six cities from a fixed seedsrc/eda_analysis.pyProduces the summaries, seasonality and correlation figuresEstablishes a working path. Claims no accuracy.
ML-Water-Quality-Classification
Four pipelines - logistic regression, random forest, XGBoost and an RBF SVM - compared under cross-validation and hyperparameter tuning on a seeded five-thousand-sample dataset.
The labels are generated, not laboratory measurements. The tracked scores measure how well each model recovers the generator's own class-correlated distributions, and say nothing about whether real water is safe to drink.
src/generate_data.pyGenerates five thousand labelled samples from a fixed seedsrc/pipeline.pyRuns all four model families through one cross-validated pipeline and tunes themPublishes tracked metrics.
Supply-Chain-Analytics-Dashboard
Cleans order data, computes operational KPIs, compares demand-forecasting baselines, and illustrates classical inventory calculations inside a single Dash application.
Versions source and notebooks only. No dataset, screenshots, or verified business metrics, so no fill-rate or forecast-accuracy claim is made.
src/data_ingestion.pyCleans the order records into a processed tablesrc/kpi_engine.pyComputes the operational indicatorssrc/demand_forecasting.pyCompares forecasting baselines, continuing if Prophet is unavailablesrc/inventory_optimizer.pyEOQ, safety stock and reorder pointssrc/supplier_scoring.pyGroups departments as an explicit proxy and scores from late deliverysrc/dashboard.pyServes all of itPublishes no metric. 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.
Academic or mathematical work, including group coursework.
ml-surrogates-thesis-data
The training and evaluation artifacts behind the thesis, in the directory layout the training scripts wrote: split scalers, test sets and loader parameters under data_created_during_training, model weights under trained_model.
Data, not analysis. 1,267 files are tracked; nineteen exceed GitHub's per-file limit and are published as release assets with a script that restores them to their real paths. Reuse of the artifacts needs prior permission, unlike the MIT-licensed upstream code.
UQ-Hydrology-Seminar-TUM
A three-person TUM seminar connecting HBV rainfall-runoff calibration, local and global sensitivity analysis, and input/output uncertainty propagation.
Group coursework. Individual ownership of each result is not claimed, and course-provided forcing data is not redistributable.
Neural-Network-Identifiability-Analysis
An educational playground for a mathematical question: if two networks agree on every input, must their parameters agree? Accompanies a TUM mathematics seminar on neural-network identification.
Source code and exploratory notebooks only. No versioned experiment configurations or numerical findings, so no empirical identifiability result is established.
src/network_isomorphisms.pyBuilds networks related by hidden-unit permutations and sign flipssrc/identifiability_checks.pyLooks for clone pairs, inactive units and sampled non-degeneracysrc/activation_analysis.pyRecords which assumptions each activation family satisfiessrc/symmetry_breaking.pyTries regularisers intended to remove the symmetrysrc/visualization.pyDraws the parameter alignmentPublishes no metric. 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.
Clearly bounded experiments. Not production systems.
Battery-SOC-Estimation-ML
Compares regression models, clustering, cycle-aware features, and a genetic-fuzzy prototype for inferring lithium-ion state of charge from voltage, current, and temperature.
A research prototype with no dataset, weights, or tracked evaluation. It must not be used to operate a battery-management system or make safety decisions.
src/data_loader.pyReads authorised NASA files, or generates discharge cycles for developmentsrc/feature_engineering.pyBuilds cycle-aware features from voltage, current, temperature and timesrc/soc_regression.pySVR, random forest, XGBoost, LightGBM and an LSTMsrc/clustering_analysis.pyK-means and Gaussian mixtures over operating regimessrc/genetic_fuzzy.pyA fuzzy estimator with its rule base optimised by a genetic algorithmPublishes no metric. 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.
Time-Series-Streamflow-Forecasting
A deterministic benchmark comparing seasonal-naive, SARIMAX, and gradient-boosted one-step streamflow predictions on a fixed-seed synthetic series.
Synthetic data. Strong scores test the evaluation pipeline and are not evidence of real-catchment validity.
Deep-Learning-Flood-Prediction-LSTM
Trains an LSTM to read thirty days of precipitation, temperature, and soil moisture and predict next-day discharge, on a deterministic rainfall-runoff generator written for the repository.
A reproducible synthetic-data demonstration. Reported metrics describe the generated benchmark only, not a validated flood-forecasting system.
src/generate_data.pyWrites the synthetic catchment from a fixed seedsrc/dataset.pyWindows it into 30-day sequences and fits the scalers on training rows onlysrc/model.pyDefines the LSTMsrc/train.pyTrains with early stopping and a learning-rate schedulesrc/evaluate.pyScores the held-out period and writes the figuresPublishes tracked metrics.
Early learning exercises, documentation and forks. Listed because the index is complete rather than curated, and not offered as evidence of engineering depth.
complete-python-warmup
An early notebook of Python, NumPy, pandas and introductory analysis exercises with their saved exploratory outputs.
Kept as a learning artifact. Several cells need external CSVs whose redistribution terms were never recorded, so the saved outputs are examples rather than independently reproducible results.
pde-problems
A command-line variant of rock-paper-scissors with input validation and a unittest suite. Despite the repository name it has nothing to do with partial differential equations.
A beginner exercise, retained under its original name rather than quietly renamed to look like something else.
git-python-basics
A minimal executable script and a standalone CSS snippet, kept from early Git practice.
Not an application and not a portfolio project. It is indexed here because the index is complete, not because it demonstrates anything.
iftaar-invitation-2026
A single-file personalised invitation for a private gathering. Guest names, animation and countdown all run in the browser; nothing is submitted, collected or measured.
An event-specific design artifact, not a reusable event-management system. The host and venue details are deliberately particular to the original invitation.
mzquadri
The GitHub profile landing page: a short statement of focus and a table of selected work in which every row carries its own evidence boundary.
Documentation. It makes no claim of its own beyond pointing at the repositories that do.
mzquadri.de
The previous portfolio landing page, retained in public so old links resolve and point at the maintained platform.
Superseded and no longer developed. It is listed so that the index does not silently omit a page that still exists.
express
A fork of the upstream Express web framework.
No authored contribution. It appears here only because this index lists every public repository, including the ones that flatter nobody.
ml-surrogates-thesis
The earlier home of the thesis codebase, archived and read-only since the work was consolidated into the fork of the upstream repository.
Superseded, and kept rather than deleted because the audited evidence cited across this site is pinned to commits here. Those links still resolve; an archived repository stays readable, and the commits do not exist in the successor.
CNN-Image-Classification-PyTorch
A compact image-classification experiment with a tracked configuration, learning history, per-class diagnostics, and an honestly recorded reference result.
A bounded educational baseline on a 15,000-image training subset. No checkpoint is versioned and no state-of-the-art result is claimed.
Insurance-Claims-Prediction-ML
Works through classification, probability calibration with Platt scaling or isotonic regression, cost-sensitive threshold selection, and SHAP-based attribution.
Source and notebooks only, with no versioned data, model, or evaluation report. It is not an underwriting, pricing, or claims-decision system.
src/data_pipeline.pyCleans, encodes and scales, fitting every transformer on training rows onlysrc/model_training.pyLogistic regression and random forest, with optional boostingsrc/calibration.pyPlatt scaling or isotonic regression, cross-validatedsrc/threshold_optimizer.pyChooses an operating threshold against a cost matrixsrc/explainability.pySHAP attributions over the fitted modelPublishes no metric. 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.
NLP-Text-Classification-Transformers
Runs TF-IDF baselines and a fine-tuned DistilBERT against the same AG News task with the same evaluation, so the two tracks can be compared rather than asserted.
Versions source and notebooks but not data, checkpoints, or metrics, so no accuracy, F1, or model-comparison claim is published.
src/data_loader.pyFetches AG News, and stops rather than silently using the offline fixturesrc/model.pyThe TF-IDF baselines and the DistilBERT configurationsrc/train.pyFits a baseline, or fine-tunes through the Hugging Face Trainersrc/evaluate.pyClassification metrics, confusion matrices and optional latencysrc/inference.pyRuns either track from a locally saved artifactPublishes no metric. Data, trained models, checkpoints, metrics, plots and run metadata are all untracked, so the repository declines to state how the two tracks compared.
Repository details are a reviewed offline snapshot. This page makes no request to GitHub, publishes no contribution counts or activity dates, and renders identically if GitHub is unavailable.
Reading the labels
Reference and experiment repositories are kept public because the reasoning in them is useful, not because they carry production evidence. Where a claim needs proof, the case study links directly to the artifact.
Research record