CAPRA — Context-Aware Privacy Risk Assessment

Five-part showcase

Research project outline

The research question, gap, and contributions in seven slides.

Framework description

The six layers of CAPRA, one at a time, culminating in a sped-up end-to-end animation.

Conceptual CAPRA demonstration

A visual walkthrough of how CAPRA connects federation, contextualisation, risk assessment, adaptation, and human oversight.

CAPRA prototype walkthrough

A practical walkthrough of the six-layer privacy risk assessment prototype for agentic AI systems.

Prototype repository

The CAPRA prototype includes n8n workflows, deterministic synthetic fixtures for healthcare, retail, and university admissions, local infrastructure, dashboards, and retained test evidence. The beta route lets a reviewer exercise the same five-stage structure without building source code.

Open on GitHub → Read the quick start

Example instantiation · university admissions

The admissions example uses twelve deterministic synthetic applicant records. A mock external-system seeder turns them into generated agent telemetry, which the CAPRA prototype processes through five executable stages over the shared Context Layer.

CAPRA reference architecture lifecycle showing Federate, Contextualise, Assess, and Adapt connected by data, context, assessment, and refinement flows, with human oversight informing adaptation — CAPRA reference architecture lifecycle. The admissions prototype instantiates this lifecycle as the five executable stages below over the shared Context Layer.

FederateNormalise generated events into telemetry_raw records with canonical event identifiers.

ContextualiseAdd actor, purpose, data, and transmission context, with immediate provenance links.

AssessConstruct scenarios and store bounded risk inferences for reviewer inspection.

RefinePrepare proposed changes with an action, target, rationale, and expected effect.

ReviewPresent an Approve or Reject form that resumes the waiting n8n execution.

A verified local-fallback window produced records through refinement. It did not submit the review form. This example demonstrates stage-level operation, not complete same-event traversal, durable decision persistence, semantic correctness, privacy effectiveness, or regulatory compliance.

The framework at a glance

Select a layer to inspect its role, the information it receives, and the output it contributes to the CAPRA assessment path.

Data Federation Layer

Observes heterogeneous telemetry, events, and PII flows from external multi-agent systems and normalises them into a common event record.

ReceivesGenerated agent events and source-system telemetry.

ContributesCanonical records with stable event identifiers.

Prototype evidencetelemetry_raw records in MongoDB.

Context Processing Layer

Enriches federated events with semantic context so later reasoning can account for actors, purposes, data categories, transmission, and domain conditions.

ReceivesCanonical telemetry and domain knowledge.

ContributesContextual assertions linked to their source event.

Prototype evidenceEnriched telemetry and prov:wasDerivedFrom links.

Risk Intelligence Layer

Constructs scenarios and produces bounded risk inferences from enriched context and the prototype's privacy-risk vocabulary.

ReceivesEnriched events, domain context, and risk concepts.

ContributesSeverity-labelled inferences with supporting rationale.

Prototype evidenceScenario, simulation, scoring, and risk_inference_results records.

Feedback & Refinement Layer

Transforms assessment outcomes into bounded refinement proposals for human consideration rather than changing the system autonomously.

ReceivesRisk results and reviewer decisions.

ContributesProposed actions, targets, rationales, and expected effects.

Prototype evidenceevaluation_results and feedback_results records.

Architecture in motion

The animation illustrates CAPRA's designed processing path from an external source through federation, contextualisation, assessment, refinement, and human review. It shows the implemented workflow structure rather than an empirically verified same-event trace across every store.

Animated illustration of CAPRA's designed processing path through its six architectural layers — The animation retains the implementation labels used when the reference workflow was captured.

Domains evaluated

Historical campaigns exercised higher-education, healthcare, and retail inputs through the same CAPRA architecture. Domain-classified records were confirmed through risk assessment, with shared refinement output in aligned observation windows. The beta route provides new deterministic synthetic fixtures for these domains.

Healthcare

Clinicians, records, treatment vs. curiosity — special-category health data.

Retail

Customer profiles, loyalty accounts, third-party sharing risk.

Student admission

Applications, transcripts, personal statements, reviewer workflows.

Try the beta prototype

Start the no-build reviewer route

Docker provisions n8n, MongoDB, Fuseki, Loki, and Grafana. Supply one generic OpenAI-compatible endpoint, or select the local llama3.2 fallback.

git clone https://github.com/gracebilliris/capra-prototype.git
cd capra-prototype
cp reviewer/env.template reviewer/.env

# Add your endpoint values to reviewer/.env, then:
./reviewer/scripts/bootstrap.sh
./reviewer/scripts/verify_route.sh
./reviewer/scripts/run_demo.sh --domain admissions --minutes 30

For the credential-free fallback, set CAPRA_LLM_PROVIDER=ollama and follow the host or container instructions in the quick start.

Open complete instructions →

What opens locally

n8nlocalhost:5679 · workflow and executions

Grafanalocalhost:3002 · privacy-risk dashboard

Fusekilocalhost:3031 · shared Context Layer

Lokilocalhost:3101 · stage logs queried through Grafana

The endpoint route is provider-neutral. Credentials stay in a git-ignored local environment file and are imported into the local n8n credential store. No browser-based credential wiring is required.

Related publications

Published peer-reviewed and preprint outputs that motivate, ground, or extend the CAPRA prototype.

Featured publication · A-ranked journal

A Federated Observability Architecture Pattern for Reliable Agentic AI Software Systems Across the AI Software Development Lifecycle

Published in Information and Software Technology, Volume 199, Article 108260 (2026). The paper presents a seven-layer architecture pattern for federating and observing heterogeneous telemetry across the AI software development lifecycle.

Read the paper View the prototype

Published

Billiris, G., & Gill, A. (2026). A Federated Observability Architecture Pattern for Reliable Agentic AI Software Systems Across the AI Software Development Lifecycle. Information and Software Technology, 199, Article 108260. https://doi.org/10.1016/j.infsof.2026.108260
Billiris, G., Gill, A., Haggag, O., Bandara, M., & Grundy, J. (2026). CPL: A Context Processing Layer for Semantic Observability in Multi-Agent AI Systems. SSRN 7194446. https://doi.org/10.2139/ssrn.7194446
Billiris, G., Gill, A., & Bandara, M. (2026). Systematic Literature Review of Data Privacy Risks in AI Systems. Science and Information Computing Conference (SAI) 2026. https://link.springer.com/book/10.1007/978-3-032-24810-7
Billiris, G., Gill, A., & Bandara, M. (2025). Privacy in the Age of AI: A Taxonomy of Data Risks. arXiv. https://doi.org/10.48550/arXiv.2510.02357
Billiris, G., Gill, A., & Bandara, M. (2025). A Taxonomy of Data Risks in AI and Quantum Computing (QAI): A Systematic Review. arXiv. https://doi.org/10.48550/arXiv.2509.20418
Billiris, G., Gill, A., Oppermann, I., & Niazi, M. (2024). Towards the Development of a Copyright Risk Checker Tool for Generative Artificial Intelligence Systems. Digital Government: Research and Practice, 5(4), Article 41. https://doi.org/10.1145/3703459
Billiris, G., & Gill, A. Q. (2024). An Initial Review of the Copyright Concerns of Generative Artificial Intelligence. ACIS 2024 Proceedings, Article 17. https://aisel.aisnet.org/acis2024/17/