Software and data

Open-source research code released together with its data, results and evaluation protocol, so that every reported figure can be regenerated and checked.

2026 · v1.0.0

FCRL-energy

Fair allocation of energy support under price shocks, with a guaranteed minimum for every region. Developed from AI with Integrity (Springer, 2026).

The repository develops the FCRL-MARL, SC-RL, XRL-URM and AC-RL frameworks of Chapter 3 into a decision system for one of the sharpest social consequences of high energy prices: how to share a fixed heating and cooling allowance among regions before need is known, so that the households least able to heat or cool their homes are protected first. Need is estimated from heating and cooling degree days, the number of residents at risk of poverty and the national energy price index; every allocation is made with the information available at the end of the previous month.

Every region receives a guaranteed minimum share of its need, protected against forecast error by adaptive conformal inference that remains valid after price shocks. The rest of the budget is shared by a policy trained to raise the worst-covered regions first while avoiding support above need, and a reserve of at most 10% is settled after each month in favour of the regions left furthest behind. Settings are selected on the 2022 energy crisis, every rule is stress-tested, and each regional allocation is explained in plain language.

  • Real data: 63 NUTS 2 regions of Greece, Bulgaria, Spain, Portugal and Italy, test seasons 2023 to 2025, the same budget for every rule
  • Coverage of the worst-off tenth of regions rose from 32% under an equal amount per resident at risk of poverty to 54% (+0.217, 95% CI 0.185 to 0.239), with no loss in total coverage
  • The least protected region rose from 24% to 41% of its need; the guaranteed minimum was met in every test region-month
  • With energy prices 30% above forecast and the budget cut by 20%, the guarantee held in 84.8% of region-months, against 72.9% under the equal amount per resident at risk of poverty
  • An interactive dashboard, republished on every run, shows every region and season and explains each allocation
Book
Springer, 2026, DOI 10.1007/978-3-032-31133-7
Archive
Zenodo, DOI 10.5281/zenodo.24033903
Data
Eurostat degree days, population, poverty rates and energy prices; synthetic demonstration
Language
Python, run automatically in GitHub Actions
Licence
MIT
Released
7 November 2026

2026 · v1.0.0

FMGCF-EWC

Vulnerability-aware early warning of heat-related health risk across European regions. Developed from Blueprints for Future Living (Springer, 2026).

The repository joins three frameworks of Chapter 6, FMGCF-EWC, MVI-HGP and FairGLA, into one runnable early warning chain. A weekly regional panel is built from Eurostat mortality and Open-Meteo temperature data; a warning model is trained by federated learning across countries, with optional client-level differential privacy and a cross-border regional graph; and a hierarchical vulnerability index separates country effects from regional vulnerability.

Warning thresholds are weighted by vulnerability only when the weighting does not lower overall detection on the validation year, and detection, false alarms and their trade-offs are reported with bootstrap confidence intervals.

  • Real data: 52 NUTS 2 regions of Greece, Italy, Spain and Portugal with complete weekly series, test year 2025
  • Test AUC 0.728; 61 of the 100 heat-related excess mortality events of summer 2025 detected with a single threshold
  • Full vulnerability weighting raised detection in the most vulnerable third of regions by 0.200 (95% CI 0.082 to 0.348) but lowered overall detection, so the validated rule kept the single threshold and reports the trade-off to decision makers
  • Federated training matched the accuracy of a centralised model (test AUC 0.728 against 0.721)
Book
Springer, 2026, DOI 10.1007/978-3-032-24178-8
Archive
Zenodo, DOI 10.5281/zenodo.24031500
Data
Eurostat weekly deaths, population and GDP; Open-Meteo temperature (CC BY 4.0)
Language
Python, run automatically in GitHub Actions
Licence
MIT
Released
7 November 2026

2026 · v1.0.0

EAA-AI

Explainable, accountability-aware AI for public-service decisions, with a tamper-evident audit ledger and human review. Developed from Empathy in Algorithms (Springer, 2026).

The repository develops the Explainable Accountability-Aware AI Architecture of Empathy in Algorithms into an auditable decision pipeline for high-risk systems under the EU AI Act. Every parameter update is written to a hash-chained ledger attributed to an accountable role, the fairness weight is calibrated to the data with 95% bootstrap confidence, and deployment is gated by six requirements: accuracy over a constant predictor, statistical parity, disparate impact, equalized odds, counterfactual consistency and the share of the explanation that rests on the protected attribute or its proxies.

A human review, recorded in the same ledger with the decision, the accountable role and a written justification, is required before release, and every run is verified automatically on GitHub Actions.

  • Synthetic public-benefit data with a strong historical bias: the unconstrained baseline is blocked (disparate impact 0.766), while the governed model passes every requirement
  • The validation results show that disparate impact and equalized odds cannot both be met closely when base rates differ, a tension the ledger records rather than hides
  • Real data, ACS public health coverage, California 2018: the governed model passes every requirement with wide margins (disparate impact 0.977, equalized odds difference 0.030)
Book
Springer, 2026, DOI 10.1007/978-3-032-32128-2
Archive
Zenodo, DOI 10.5281/zenodo.24019871
Data
Synthetic public-benefit data; American Community Survey PUMS 2018, California, via Folktables
Language
Python, run automatically in GitHub Actions
Licence
MIT
Released
6 November 2026

2026 · v1.0

FMA-AI

Federated multi-agent AI for privacy-preserving, energy-aware control of public transport, evaluated on the MBTA GTFS feed.

Route-level agents learn a shared speed-regulation policy through synchronous federated averaging (FedAvg), so that operating data remain on the route that produces them. Each agent minimises traction, auxiliary and congestion-related energy, subject to a bounded loss in on-time performance. The framework is evaluated in a simulation built on the public GTFS feed of the Massachusetts Bay Transportation Authority and compared with a centralised optimiser and with route-isolated learners.

Across 20 random seeds, with six training routes and 24 hold-out routes, the federated policy reduced energy use and CO₂ emissions by 0.78 ± 0.07 % on the training routes and by 0.88 ± 0.06 % on routes it had never seen (mean ± 95 % confidence interval, with an on-time tolerance of two percentage points). Route-isolated learners produce no model for unseen routes, which is the practical argument for federation.

The October 2026 revision supersedes the figures posted in October 2025: the implementation was corrected and every result regenerated, and the repository documents each correction. An independent replication in GNU Octave reproduces the route selection exactly and yields overlapping confidence intervals.

Paper
SSRN, revised October 2026
Archive
Zenodo, DOI 10.5281/zenodo.23242130
Data
MBTA GTFS feed, Summer 2025, redistributed under the MassDOT Developers License Agreement
Languages
MATLAB, GNU Octave, Python
Licence
MIT (code)

2026 · v0.1.0

AegisOne Guardian

Leakage-resistant temporal evaluation of a static malware classifier, with an EMBER2024 benchmark.

The repository contains the evaluation protocol, feature extractor and tooling behind the static file classifier of AegisOne, an endpoint-protection prototype for Windows. It is published as a research artefact; the product itself is not included. Its contribution is methodological. Random train and test splits, accuracy or AUC, and point estimates of the false-positive rate each tend to flatter a classifier, and the repository implements the stricter alternative end to end.

  • Temporal splits only, with SHA-256 checks that no binary appears in two partitions
  • Thresholds chosen on validation data and reported once on a later test set
  • Detection of unseen malware families and false positives on unseen publishers reported separately
  • False-positive rates given as one-sided 95 % Clopper–Pearson upper bounds
  • Calibration reported through the Brier score, expected calibration error and a reliability table
  • One 48-feature schema implemented in Python and C#, with parity verified in continuous integration

A public benchmark applies the protocol to EMBER2024 and asks whether a threshold set to deliver a 1 % or 0.1 % false-positive rate still delivers that rate on later weeks. Results on the AegisOne corpus itself have not yet been published.

Benchmark
EMBER2024 (.NET subset), run automatically in GitHub Actions
Languages
Python, C# (.NET 8)
Licence
MIT
Released
13 October 2026

2026 · v1.0.0

MATLAB for Smart Infrastructure

Companion code for Chapter 15 of MATLAB for Civil Engineers (Springer, 2025): five reproducible case studies in digital twins, structural health monitoring, IoT and optimisation.

The repository turns five case studies from the closing chapter of the book into self-contained scripts that fix their random seeds and export every figure and table they report. The models and parameters follow the printed text. For public release, validation and verification were added where the book reports results on the training data alone.

  • Digital twin of a bridge: bagged regression trees predict stress from streaming sensor data, validated on a hold-out set, with threshold alerts
  • Structural health classification: bagged trees and a shallow neural network, assessed with leave-one-out cross-validation
  • IoT structural monitoring: strain, tilt and vibration streams, with persistence-based alarms that suppress isolated exceedances
  • Traffic signals in Athens: green times at five locations optimised with a genetic algorithm and checked against the closed-form optimum
  • Wastewater treatment plant: a Pareto front between cost and environmental impact under a capacity constraint, with a compromise design
Book
Springer, 2025, DOI 10.1007/978-3-031-84673-1
Archive
Zenodo, DOI 10.5281/zenodo.23555622
Language
MATLAB R2023b or later
Toolboxes
Statistics and Machine Learning, Optimization, Global Optimization
Licence
MIT
Released
22 October 2026

How to cite

All six repositories include a CITATION.cff file; BibTeX entries are in the full bibliography.

2026

Sargiotis, D. (2026). FCRL-energy: Fair and sustainable allocation of energy support under price shocks (Version v1.0.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.24033903

2026

Sargiotis, D. (2026). FMGCF-EWC: An operational, vulnerability-aware early warning chain for climate-related health risk (Version v1.0.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.24031500

2026

Sargiotis, D. (2026). EAA-AI: Reference implementation of the Explainable Accountability-Aware AI Architecture from Empathy in Algorithms (Version v1.0.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.24019871

2026

Sargiotis, D. (2026). FMA-AI: Federated multi-agent AI for energy-aware public transport (MBTA GTFS) (Version v1.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.23242130

2026

Sargiotis, D. (2026). AegisOne Guardian: Leakage-resistant evaluation of a static malware classifier (Version 0.1.0) [Computer software]. GitHub. https://github.com/uomoathens/aegisone-guardian-eval

2026

Sargiotis, D. (2026). MATLAB for Smart Infrastructure: companion code for Chapter 15 of MATLAB for Civil Engineers (Version v1.0.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.23555622

↑ ↓ to moveEnter to openEsc to close