I study the ethics of AI-assisted decisions as a question of evidence: what changes when a system changes, who bears the consequences, and who signs off.
My doctoral research asked how heterogeneous biomedical evidence could support explainable inference. I carry that concern into AI ethics today: what warrants a decision, what its evaluation leaves unseen, and who can challenge it.
PhD · Computer Science · University of Colorado Boulder
I investigated drug–drug interaction characterization through statistical inference and Semantic Web technologies. This established my foundation in connecting heterogeneous evidence to an explainable inference.
Characterization of Drug-Drug Interactions through Statistical Inference and Semantic Web Technologies
Peer-reviewed article · JAMIA · online 2016, issue 2017
D3 integrated biomedical knowledge and inference rules to propose explanations for drug interactions. The question carried forward is how a prediction can be supported by an inspectable chain of evidence.
Drug-drug interaction discovery and demystification using Semantic Web technologies
bioRxiv preprint · not yet peer reviewed
D4 connected deep-learning predictions with mechanistic inference using D3. It extended the effort to examine what lies behind a predicted interaction.
D4: Deep Drug-drug interaction Discovery and Demystification
Peer-reviewed article · Journal of Clinical Pharmacology
My ontology-enhanced DDI study combines structured biomedical knowledge with deep learning for mechanistic prediction. It continues the work on making the basis of an inference interpretable.
Ontology-Enhanced Deep Learning for Mechanistic Prediction of Drug-Drug Interactions: A Clinically Interpretable Framework
Current research · empirical and normative questions
I now examine what evaluation permits us to conclude about a decision, whose outcomes it conceals and who remains responsible. RIDI contributes to the evaluation question; iSCARB contributes to human-owned judgment. Together they support a wider agenda in justice, accountability and delegation.
RIDI manuscript and iSCARB preprint: current work with separate research status
The early work is biomedical informatics. The connection to AI ethics is the research direction I develop from that foundation; these publications retain their original scope and status.
I connect empirical audits of AI behaviour with questions about legitimate decisions. The studies below are directions for collaboration, alongside the existing work linked in each pillar.
Who experiences a different outcome when systems receive the same evaluation score?
I examine how model updates, rankings and selection rules change who or what is chosen. I treat these changes as a reason to investigate their justification and distribution; change alone does not establish unfairness.
Compare equivalent-performing systems at the level of affected individuals, then examine which changes are necessary, avoidable or unjustified.
What must be recorded so that a person can question an AI-assisted decision?
I connect the representation and provenance of evidence to the reasons offered for a decision. The goal is to make its assumptions, missing knowledge and limits open to scrutiny.
Trace how changes in source data, terminology or evidence alter a recommendation, and evaluate whether the resulting explanation supports meaningful review.
Which judgments may an institution delegate to AI, and which responsibilities must people retain?
I study how people can own, defend and revise a judgment when AI produces part of the answer. Oversight must involve the ability and authority to intervene.
Evaluate when people can detect a system’s limits, challenge its advice and revise their decisions under changed conditions.
RIDI measures what an unchanged evaluation score leaves unseen. In a preregistered, open experiment, five standard retrieval metrics stayed identical while correctness changed in 17.27% of 800 queries, far above the repeat-execution floor. Regulators now ask for pre-specified acceptance criteria before an AI update; I use this work to ask what such a criterion must record before anyone signs off.
Manuscript · submission-ready, not yet peer reviewed
In a preregistered retrieval experiment, five standard retrieval metrics were identical, yet benchmark-defined answer correctness changed for 17.27% of 800 queries. Registered extensions, an exact control frontier and a production update test the limits of that result. Submission-ready; not yet peer reviewed, accepted or published.
The current experiment concerns retrieval and answer correctness. It does not by itself demonstrate discrimination, clinical benefit or the fairness of an institutional decision.
Published articles, preprints, manuscripts and open data are identified separately. The full publications page covers my wider collaborative record.
Peer-reviewed article · 2026-06
A clinically interpretable DDI framework combining ontology-based embeddings with deep learning to predict pharmacokinetic and pharmacodynamic interaction mechanisms.
Manuscript · submission-ready, not yet peer reviewed · 2026-09-26
In a preregistered retrieval experiment, five standard retrieval metrics were identical, yet benchmark-defined answer correctness changed for 17.27% of 800 queries. Registered extensions, an exact control frontier and a production update test the limits of that result. Submission-ready; not yet peer reviewed, accepted or published.
Open data · CC BY 4.0 · 2026-09-26
Source-data tables for the main and Extended Data figures, registered-extension summaries, P5b practitioner-equivalence outputs and second-judge robustness summaries, so the reported results can be audited. Third-party benchmark corpora are not redistributed.
Preprint · not yet peer reviewed · 2026-09-25
A framework and nine-chapter implementation for human-owned engineering judgment in systems containing AI. It combines evidence, explicit limits, changed conditions and justified revision; learning effectiveness remains an evaluation question.
Peer-reviewed article · 2026-08-28
With Redhwan Nour, this study evaluates selective, context-aware AI assistance for chest X-ray interpretation rather than assuming AI should always be shown.
Research statuses as recorded on 26 September 2026; linked records provide the source.
All publications & profiles →These projects provide contexts for future ethics research and co-development. Each has its own development and validation stage.
Founder, HEALTHx — clinical innovation platform; invention disclosure under review. A development context for studying how clinical knowledge, protocols and AI-supported workflows can be documented, evaluated and governed with hospital partners. Partner and clinical validation remain work to establish.
Explore HEALTHx →A working prototype for organizational and workforce decisions. It offers a context for studying how categories, evaluation rules and approval authority shape consequential choices.
Explore MIYAR →I welcome collaboration across computer science, philosophy, medicine, law and public policy. Start with a decision problem, the evidence available and a question we can investigate together.
Joint empirical and conceptual studies, independent replication, evaluation design and co-supervision. Bring a method or disciplinary perspective that helps test the question.
Academic collaboration →Thesis and project questions in AI ethics, evaluation, evidence provenance and human judgment. A starting proposal should identify a decision, a feasible method and the evidence needed.
Research opportunities →Co-develop research around a real decision process: domain expertise, appropriate data access, evaluation criteria and clear human responsibilities.
Discuss a partnership →Selection-change auditing and the exact identity–utility frontier for ranking updates.
Strategy, organizational design, job evaluation, workforce planning, compensation and approval workflows.
Biomedical benchmark research and reproducibility resources.
Historical research repository. The associated 2022 article was retracted in 2023; it is not used here as validated evidence.
Publisher status notice →Course resources and educational systems.