Archimedes Unit, Athena Research Center
Ariadne AI Lab. Unraveling the thread of complex government procedures

Research

AI development projects with Greek public agencies. Each one is built on public or partner data, evaluated by domain experts, and operated under human oversight.

Projects

Multilevel Governance

Partner: Ministry of Interior

Status: Phase A delivered, final evaluation round in progress. Phase B planned.

The competences of an agency are scattered across its founding act and every act that amended it. Codification by hand for the whole registry is not feasible, and a consolidated, verified record is the basis for allocating responsibilities across levels of government.

Multilevel Governance codifies the competences of Greek public agencies for the information system of the Ministry of Interior that records which unit exercises which competence. The lab built the pipeline that performs the codification. It reads the founding act of an agency from the Government Gazette (FEK) or its statute from the GEMI business registry, builds the chain of amending acts, extracts the competences of each organizational unit, applies every amendment in chronological order, and exports the consolidated text with the list of amending provisions. Deterministic rules and a language model share the work, and each applied change is validated. The staff of each agency verify the output.

Evaluation on 29 agencies against a ministry gold set and expert corrections gives 89% recall on competences and 84% exact match on unchanged text. The deliverable covers the 2,313 agencies of the registry in four formats. Phase B adds targeted updates from a single new amending act and a review interface for the history of amendments.

Prosontologio

Partner: Ministry of Interior

Status: Evaluation phase.

New legislation can change the requirements for a position without naming the decree it affects. Automatic detection keeps the register of qualifications current, and independent evaluation gives a measure of how reliably such changes are found.

Prosontologio concerns the qualification requirements of public sector positions under p.d. 85/2022 and the legislation that changes them. The lab defined the use cases, built the dataset of 85 Government Gazette issues plus a hidden test set of 48, and independently evaluates an application that detects legislation changing qualification requirements. Research questions include structure extraction from long Greek legal documents, amendment resolution, and evaluation of rationale quality.

RegTech

Partner: Hellenic Data Protection Authority, Ministry of Digital Governance

Status: Phase A on public data. Later phases on historical and live cases.

Complaint handling at a regulatory authority depends on finding the applicable legislation, decisions, and guidelines and on drafting reasoning that cites them. An assistant that performs this part shortens handling time while the officer keeps the decision, and the same design transfers to other authorities.

RegTech is an assistant for the case officers of the Hellenic Data Protection Authority, developed with the Ministry of Digital Governance. It classifies a case by thematic category, retrieves the applicable sources (legislation, Authority decisions, EDPB guidelines, CJEU case law) from a curated knowledge base of about 3,000 documents, and drafts the legal reasoning section by section with a citation for every claim. The officer approves or rejects each section. The system cites only sources in the knowledge base and declares when a question is beyond its coverage.

Phase A runs on public data, about 2,000 published decisions, with weekly expert evaluation and a leaderboard of open and closed models per task. Research questions include grounded legal drafting with abstention, retrieval over hierarchical legal sources, and measuring usefulness through expert feedback. The design is model-agnostic.

RegTech demo →

Building Permits

Partner: Ministry of Environment and Energy, Technical Chamber of Greece

Status: Starting.

Permit review is slow and uneven because the required documents depend on the location, use, and type of works, and because an inspector meets each file without context. Tools that derive the requirements from the conditions of the project, verify what is submitted, and summarize a file shorten review and make it consistent across offices.

Building Permits brings AI tools into e-Adeies, the national building permit system, in view of the new Building Control Centers and the independent permit inspectors. The tools check the completeness and correctness of submitted documents against the applicable rules, prefill building terms and approvals with verification of site plans and coverage diagrams, and summarize a file for an inspector who sees it for the first time, with conflict-of-interest screening before assignment.

The first proof of concept produces a personalized list of the supporting documents, approvals, and studies that a permit application requires. An AI component proposes the type of works from the free-text description, geospatial queries detect the conditions of the project, and a deterministic component maps the conditions to the required documents, with the reason and the deciding component recorded for each. Diagram verification follows, drawing on a large archive of coverage diagrams with inspection notes. Research questions include combining rules, geospatial queries, and vision-language models for technical drawings, and measuring accuracy against human inspection.

Knowledge Base Builder

Partner: Ministry of Digital Governance

Status: Starting.

Every agency that wants an assistant over its own texts otherwise needs its own application, infrastructure, and maintenance. One builder, run once per topic and agency, reduces that work to the collection, curation, and structuring of texts, and every answer carries a reference that can be checked, so answers stay consistent across staff and agencies.

The Knowledge Base Builder turns the official texts of a public agency into a structured knowledge base file that plugs into the commercial AI assistant subscription the agency already holds, with no new application, server, or platform. The knowledge base holds the texts of official sources, segmented by article and paragraph, with the provenance of every item. Staff ask in Greek and receive the provision itself, with the article reference and a link to the official source. The assistant returns verbatim text instead of a summary, ranks binding provisions above guidance and studies, and repeats the search when the results do not contain the act itself.

Sources are limited to the National Printing Office, the EU Publications Office, Diavgeia, official guidance, and the documents of the agency. The pilot at the Ministry of Digital Governance spends two months on collection, curation, base construction, and integration, and one month on legal review of the answers and pilot use in one directorate with real user questions.

Resources

Datasets, benchmarks, models, and code released by the lab will be listed here.

Publications

Publications of the lab will be listed here.