A Unified Framework for Explainable Front-to-Back AI Pipelines

Thesis Abstract

This study addresses the escalating opacity of end-to-end AI systems by proposing a unified framework that integrates explainability across front-end data preprocessing, model inference, and back-end deployment while preserving performance. The aim is to develop and validate an interpretable, auditable pipeline that enables stakeholder-friendly explanations at every stage of the AI lifecycle, thereby supporting responsible decision-making and regulatory compliance. Specific objectives are (1) to conceptualize a holistic explainability framework that links data provenance, feature attribution, model-agnostic explanations, and deployment-level transparency; (2) to operationalize the framework into a modular front-to-back architecture with standardized interfaces for explanation generation and provenance capture; (3) to empirically evaluate the impact of the unified framework on model trust, debugging efficiency, and decision accountability; (4) to compare explainability pathways across three domains—healthcare, finance, and autonomous systems—to identify domain-specific requirements and trade-offs; and (5) to formulate design guidelines and governance checklists for practitioners and regulators. The research adopts a mixed-methods design grounded in the postulates of the Theory of Explanatory Coherence and the Causal Misalignment framework. A multi-site, cross-domain study will be conducted, comprising quantitative experiments with a total sample of 900 real-world records (300 per domain healthcare, finance, autonomous systems) curated from partner institutions and public benchmarks, supplemented by qualitative interviews with 40 domain experts. Data collection instruments include (i) standardized explainability dashboards measuring feature attribution fidelity, counterfactual plausibility, and provenance traceability; (ii) performance metrics for core models (accuracy, calibration, latency) to assess trade-offs; (iii) a structured interview protocol to capture practitioner perceptions of trust, interpretability, and governance needs; and (iv) a usability questionnaire to evaluate the practicality of the proposed interfaces. Validity and reliability are ensured through triangulation, pilot testing (n=60) of instruments, and inter-rater reliability checks (Cohen’s kappa > 0.75) for qualitative coding. Data analysis will employ regression analyses to quantify the relationship between explainability features and trust metrics, multivariate ANOVA to compare effects across domains, and thematic analysis to extract nuanced insights from expert interviews. A formalized analytical framework will be used to map data provenance to lineage-aware explanations, while model-agnostic interpretation tools (SHAP, LIME) are integrated with provenance metadata to generate cohesive front-to-back narratives. The expected findings indicate that a unified front-to-back explainability framework enhances stakeholder trust without sacrificing predictive performance, with explainability gains most pronounced in healthcare and autonomous systems due to higher safety and accountability requirements. It is anticipated that provenance-aware explanations will improve debugging efficiency by reducing time-to-trace for data quality issues and model drift, and that domain-specific customization (e.g., causal explanations in healthcare, risk-focused narratives in finance) will be essential for practical adoption. The study also expects to identify a principled balance point in the trade-offs between model complexity, explanation fidelity, and system latency, offering a decision-support model for selecting appropriate explanation techniques under regulatory constraints such as GDPR and sector-specific guidance. Contributions to knowledge include (1) a concrete, testable unified architecture that structurally integrates data provenance, interpretable modeling, and deployment transparency; (2) an operationalized methodology to generate coherent explanations across the entire AI pipeline; (3) empirical evidence on the impact of end-to-end explainability on trust, accountability, and efficiency across multiple domains; and (4) policy-relevant governance guidelines and a runtime dashboard design that facilitate reproducible audits. The main conclusion is that end-to-end explainability is feasible and beneficial when embedded as a modular, provenance-aware framework that aligns explanation strategies with domain requirements. Recommendations emphasize standardization of explanation interfaces, adoption of provenance schemas, investment in domain-adaptive explanation libraries, and continuous alignment with evolving regulatory standards.