When most journalists write about Laura Kövesi, they focus on palace intrigue, political backlash. And high-profile indictments. Those stories matter, but they miss the operational reality underneath her work. The institution she runs-the European Public Prosecutor's Office (EPPO)-is also one of the most ambitious cross-border platform engineering projects in the public sector today. It has to move sensitive evidence across jurisdictions, languages, legal systems. And legacy IT estates while staying inside strict privacy and due-process rules.
The hardest part of fighting EU fraud isn't the courtroom argument; it's the data architecture that lets evidence cross borders without crossing trust boundaries.
In this post we look at the engineering lessons hidden inside Laura Kövesi's mandate. We will treat EPPO as a distributed system, examine how digital evidence pipelines work. And extract concrete patterns for building regulated software platforms. If you design case-management systems, legal-tech products, fintech compliance stacks. Or any platform that must prove what happened and who touched what, there's a lot to learn here. Read our guide to GDPR-compliant data pipelines Explore SRE observability for regulated workloads
Who Is Laura Kövesi and Why Engineers Should Care
Laura Kövesi is a Romanian prosecutor who first became widely known as the head of Romania's national Anticorruption Directorate (DNA). Between 2013 and 2018 she built a reputation for pursuing complex financial-crime cases against politically connected defendants, often using digital trails, wiretaps. And financial records as primary evidence. Her office's conviction rates were unusually high for the region, in part because the DNA invested early in case-management tooling and data-driven investigations. That background is directly relevant to her current role.
In 2019, Laura Kövesi was appointed the first european Chief Prosecutor of the European Public Prosecutor's OfficeThe EPPO became fully operational on 1 June 2021 with prosecutors in 22 participating EU Member States. Its job is to investigate and prosecute crimes against the EU budget-VAT carousel fraud, corruption, money laundering. And misuse of EU funds. The office is novel because it's the first supranational prosecution body in the EU with the power to bring cases in national courts.
From an engineering perspective, Laura Kövesi inherited a greenfield organization that had to deliver software workflows before it could deliver convictions. The EPPO couldn't simply nationalize an existing system; it had to create a shared operating model on top of sovereign legal infrastructures that's the same challenge faced by teams building global compliance platforms, health-data networks. Or cross-border payment systems: unify process without centralizing authority.
The EPPO as a Distributed Software Platform
The cleanest mental model for EPPO is a distributed platform with a central orchestrator and local worker nodes. The central office in Luxembourg sets policy, coordinates major cases. And manages the case-management system. The "worker nodes" are the European Delegated Prosecutors (EDPs) located in each participating country. Each EDP operates inside its own legal and technical environment, much like a regional microservice that owns its own data store but must participate in a larger workflow.
That distribution isn't optional. National procedural law governs how evidence is collected, how suspects are questioned. And how trials are run. The EPPO can't override those rules. Instead, it abstracts the common elements-PIF crimes, evidence standards, mutual legal-assistance requests-into a shared layer. In software terms, this is the difference between a monolith that forces every country into one runtime and a microservices architecture that lets each domain keep its own bounded context.
Event-driven patterns are essential here. A case opened in Portugal may need bank records from Germany, witness testimony from Italy. And asset freezes in France. Each hand-off is an event, and if the message bus drops a request,Or if two EDPs update the same record without conflict resolution, the case can stall or collapse. Patterns like idempotency keys, outbox tables, and compensating transactions aren't academic; they're the difference between a conviction and a dismissed charge.
Cross-Border Evidence Pipelines and Data Engineering
Modern fraud leaves a digital exhaust: SWIFT messages - VAT returns, procurement PDFs, encrypted chat logs, GPS tracks, AIS maritime data, and blockchain transactions. The job of an EPPO investigation is to ingest that data - normalize it, enrich it. And make it searchable across languages and formats that's a data-engineering problem first and a legal problem second.
In production environments, we have found that cross-border pipelines fail not at the algorithm layer but at the schema boundary. One country may call a suspect field "persoana_fizica," another "natural_person," and a third "PersonnePhysique. " Without a canonical data model and explicit mapping rules, joins break silently. A sensible architecture uses an ELT pipeline: land raw evidence in object storage, apply typed transformations. And publish canonical events to a stream such as Apache Kafka. Each jurisdiction consumes events it's authorized to see.
Provenance must be embedded from the first byte. Hash the file at ingestion using SHA-256, sign metadata with XAdES or PAdES. And bind timestamps to a trusted authority via RFC 3161 Internet X. And 509 Public Key Infrastructure Time-Stamp ProtocolIf a defense lawyer later claims the evidence was tampered with, the prosecution must produce an unbroken audit trail. Data engineering without integrity guarantees is just a faster way to create inadmissible evidence,
Digital Forensics, e-Evidence,And Chain-of-Custody Systems
Chain of custody is an audit log with legal consequences. Every copy, transfer, analysis. And disclosure of a digital artifact must be recorded in a way that proves the artifact hasn't changed. That means write-blockers for disk imaging, forensic containers for packaging. And immutable storage for the master copy. Tools like Autopsy, Sleuth Kit, and Volatility are common in labs, but the storage and workflow layer is just as important.
Object-lock policies and versioned storage can approximate WORM (write-once-read-many) behavior in cloud environments. Each access should generate an append-only log entry that includes who requested access, why, what was returned. And which hash was verified. Ideally the audit log itself is tamper-evident; Merkle-tree-based logs or append-only ledgers make it computationally infeasible to rewrite history without detection. Laura Kövesi's office can't afford a "we think the log is intact" moment in court.
The EU's proposed e-Evidence package adds another layer of complexity. It would allow judicial authorities in one state to directly request preservation or production of electronic evidence from service providers in another. That sounds like a standard API integration. But it requires mutual trust in identity, lawful authority - data minimization. And timing. A malformed request could disclose private messages to the wrong jurisdiction or miss a critical window because a preservation order was not machine-readable.
AI and Machine Learning in Anti-Corruption Investigations
There is no way to manually review millions of invoices, contracts. And transactions. Machine learning can flag anomalies: a company that wins 90 percent of tenders in a small region, a VAT refund pattern that traces a circular goods movement. Or a network of shell companies sharing a single postal address. Graph databases such as Neo4j or Amazon Neptune are especially useful for mapping relationships between entities, bank accounts. And transactions.
But AI in prosecution is high risk. A model that surfaces a suspect because of biased training data can violate fundamental rights. Under GDPR Article 22, individuals have rights related to automated decision-making. And judges are generally hostile to black-box evidence. Any ML pipeline used in this context needs explainability layers-SHAP values, LIME, or attention maps-plus model cards that document training data, performance metrics, and known limitations. Human prosecutors must remain in the loop; the algorithm should recommend, not decide.
The practical pattern is to treat ML as a triage tool, not a verdict engine. An anomaly score sends a file to the top of a human reviewer's queue. The reviewer can then request lawful evidence collection. This keeps the system efficient while preserving the procedural safeguards that Laura Kövesi's office depends on. Learn about responsible AI patterns for regulated industries
Identity, Access Control, and Judicial Trust Boundaries
Before two prosecutors can share evidence, they must be certain of each other's identities and authority. EPPO can't use a simple corporate SSO. It needs high-assurance identity backed by eIDAS-notified electronic identities, smart cards. Or hardware security keys, federated through protocols such as OpenID Connect (OIDC) or SAML 2. Multi-factor authentication is table stakes; phishing a prosecutor's credentials could expose active investigations or reveal witnesses.
Authorization is even harder. A European Delegated Prosecutor in Greece shouldn't browse cases in the Netherlands by default. Access must be attribute-based: role, case assignment, jurisdiction, clearance, and legal basis. Policy engines like Open Policy Agent (OPA) can evaluate these attributes at request time. But the attribute sources themselves must be authoritative. If a national HR system says someone is still employed when they have actually left, the policy is worthless.
This is a zero-trust scenario. Every access request is verified, sessions are short-lived, and lateral movement is constrained by network segmentation. Audit logs feed a SIEM that can detect impossible travel, off-hours access. Or attempts to export bulk data. Identity isn't just a login gate; it's the trust boundary that holds the entire distributed platform together.
Observability, Auditability. And Compliance Automation for Prosecutors
Observability for a prosecution platform is different from observability for a consumer app. Uptime matters, but so does the ability to answer historical questions: who opened this file, when was it copied,? Where was it sent,? And what decisions were made along the way? That requires structured logs, trace IDs that span cross-border requests. And metrics that measure process integrity as well as throughput.
Instrumentation should use open standards such as OpenTelemetry so that national systems and central systems can correlate events without vendor lock-in. Logs must be append-only and separated from operational logs to prevent tampering. A SIEM like Splunk, Elastic. Or a sovereign equivalent becomes the nervous system of compliance. Dashboards should highlight not just failed API calls but also anomalous access patterns and retention-policy exceptions.
Compliance automation reduces the chance that a case is compromised by an expired retention rule or an overlooked data-subject request. For example, when a suspect is acquitted, national law may require deletion or sealing of certain data. A manual spreadsheet isn't good enough. Engineering teams should implement policy-as-code: retention labels, legal-hold flags. And deletion workflows that are version-controlled, tested. And audited. The same discipline applies to any platform that handles health records, financial transactions. Or personal data.
Interoperability Challenges Between National and EU Systems
The EPPO doesn't operate in a vacuum. It must interoperate with Eurojust, Europol, OLAF, national judicial systems, and frameworks such as e-CODEX, the European Investigation Order (EIO), and the European Arrest Warrant (EAW). Many of these systems were built at different times, using different standards: SOAP/XML, REST/JSON, EDIFACT. Or bespoke file drops that's the enterprise-integration challenge every large organization faces, just with higher stakes.
The right architectural response is an adapter or gateway pattern. A central API gateway accepts canonical EPPO requests and translates them into the formats expected by each partner system. Transformation layers handle schema differences, character encodings, and idiosyncratic national procedures. Message queues buffer requests when a partner system is down. And idempotency prevents duplicate arrest warrants or preservation orders.
Testing these integrations is hard because you can't use real case data in a staging environment. Synthetic data generators, conformance test suites, and bilateral sandboxes are essential. And the EU's Council Regulation (EU) 2017/1939 defines the legal framework, but the technical interoperability is built incrementally through working groups, pilots, and painful production incidents that's normal. The lesson is to budget for integration time the way you budget for feature development.
Practical Engineering Takeaways From the Laura Kövesi Model
There are three patterns any engineering team can borrow from the EPPO model. First, use domain-driven design. Legal concepts such as case, suspect, evidence. And indictment should be modeled explicitly in code, not flattened into generic tables. Bounded contexts protect each jurisdiction's rules while shared kernels handle cross-border concepts. When the domain is complex, the code must reflect that complexity or it will leak into bugs.
Second, build compliance and auditability in from day one, not as a bolt-on. Immutable logs - signed artifacts, access policies. And retention workflows should be part of the initial architecture. Retrofitting them later is expensive and error-prone. In regulated environments, compliance is a feature, and it deserves the same test coverage and code review as any business logic.
Third, assume heterogeneity and design for it. National legal systems, like enterprise legacy systems, won't unify overnight. Use strangler-fig patterns to modernize components, API gateways to bridge protocols, and event-driven choreography to avoid tight coupling. Laura Kövesi's platform succeeds when it respects local autonomy while enabling cross-border coordination. The same principle applies to multinational SaaS products, banking consortia. And healthcare data networks.
Frequently Asked Questions About Laura Kövesi and Technology
Who is Laura Kövesi?
Laura Kövesi is a Romanian prosecutor and the first European Chief Prosecutor of the European Public Prosecutor's Office (EPPO). She previously led Romania's National Anticorruption Directorate and is known for using data-driven methods in complex financial-crime investigations.
What is the EPPO and why is it a technology story?
The EPPO is the EU's supranational prosecution body for crimes against the EU budget it's a technology story because it must coordinate prosecutors across multiple legal systems, move digital evidence across borders. And maintain strict chain-of-custody and privacy controls at scale.
How does EPPO handle digital evidence across borders?
It relies on legal instruments such as the European Investigation Order and proposed e-Evidence rules, supported by case-management systems, secure communication networks, and data pipelines that preserve provenance through hashes, digital signatures, and timestamping.
Can AI replace prosecutors in anti-corruption work?
No. AI can triage large datasets and surface anomalies, but prosecution requires human judgment, legal authority. And due process. Machine-learning outputs must be explainable and reviewed by qualified investigators.
What can private-sector engineering teams learn from Laura Kövesi's work?
Teams can learn to treat compliance as architecture, design for heterogeneous environments, embed auditability from the start. And keep humans accountable when algorithms support high-stakes decisions.
Conclusion: Building Trust at Scale Is an Engineering Problem
Laura Kövesi's public profile is built on courtroom outcomes. But the work behind those outcomes is deeply technical. The EPPO is a case study in how to build a distributed, regulated platform that must prove its own integrity to courts, citizens, and partner governments. Every design choice-from identity federation to immutable logs to schema mapping-has a direct impact on whether justice is delivered or delayed.
For software engineers, the lesson is broader. Whether you're building legal tech, fintech compliance, health data exchanges, or cross-border logistics platforms, you will face the same forces: fragmentation, sovereignty, audit requirements. And the need for human accountability. Start with domain-driven design, instrument everything. And never let the data pipeline outrun the trust model. If you want help architecting a regulated platform, schedule a technical architecture review with our team.
What do you think?
Should cross-border judicial platforms adopt open-source interoperability standards like OpenID Connect and OpenTelemetry,? Or do sovereign legal systems require fully custom protocols?
How should engineering teams balance the speed of AI-assisted fraud detection with the explainability required by courts and data-protection law?
Is immutable, tamper-evident logging now a baseline requirement for any platform that handles evidence, financial records,? Or sensitive personal data?
