Retrieval is not the last risk
Even with relevant context, generated prose can loosen obligations, attach unsupported citations, mix revisions, or invent organization-defined values.
My dissertation studies a narrow failure in compliance RAG: relevant evidence can be retrieved, yet a model-written final answer can still introduce unsupported compliance content.
ComplianceGPT is a citation-contract answerer for audit-facing QA over NIST SP 800-53 Rev. 4 and Rev. 5. The language model selects evidence IDs from a revision-scoped candidate pool; it does not write the evidence-bearing final answer. A deterministic orchestrator assembles the answer from canonical clauses, approved organization-profile values, or explicit unresolved states, then a rule-based verifier checks the released artifact.
Even with relevant context, generated prose can loosen obligations, attach unsupported citations, mix revisions, or invent organization-defined values.
S7 retrieval builds a clause pool. An LLM selects source IDs. The Contract Orchestrator resolves them against the Canonical Clause Store and ODP policy; the Runtime Verifier checks citation linkage, revision scope, source spans, status, and ODP behavior.
Under matched S7 retrieval, strict answer-pass rose from 0.22 to 0.64 on Rev. 5 and 0.75 on Rev. 4. In no-profile ODP cases, the baseline false-completed 60/63 and 19/19; ComplianceGPT made no false-complete ODP insertions in those evaluated subsets.
ComplianceGPT does not determine whether an organization is compliant and does not prove selected evidence is complete. Its claim is narrower: conditioned on selected source IDs and approved profile values, final-answer realization should not introduce unsupported compliance content.
I teach computer science through systems students can build, test, debug, explain, and defend.
Courses I am prepared to teach across applied AI, cybersecurity, and software systems.
Introduction to Artificial Intelligence, Applied Generative AI / LLM Systems, and Machine Learning Applications.
Security in Computing, Software Vulnerabilities and Security, and Cybersecurity Compliance and Governance.
Software Engineering, Web Application Development, Database Systems, Data Structures, and Programming I / II.
My classroom and mentoring work spans security, software vulnerabilities, capstone engineering, architecture review, technical presentations, and interdisciplinary software projects.
Teaching assistant for CEN 5013 and CEN 5079, Software Vulnerabilities and Security, at Florida International University.
Mentors teams through architecture, implementation, testing, documentation, technical presentations, and final demonstrations.
Evaluates interdisciplinary capstone projects for system design, technical execution, communication, and real-world usefulness.
Peer-reviewed and co-authored work across compliance, security, Zero Trust, applied AI, satellite imagery, and health informatics.
Jennifer Bolton, Wenjia Wang, Catherine Coccia, Vijaya Narayanan, Marcia Varella, Jean Hannan, Cristina Palacios. Maternal and Child Health Journal, 1–11. Springer US.
Wenjia Wang, Seyed Masoud Sadjadi, Naphtali Rishe, Arpan Mahara. 2025 IEEE Conference on Communications and Network Security (CNS), 1–6. IEEE.
Arpan Mahara, Md Rezaul Karim Khan, Liangdong Deng, Naphtali Rishe, Wenjia Wang, Seyed Masoud Sadjadi. Applied Sciences, 15(3), 1027. MDPI.
Arpan Mahara, Md Rezaul Karim Khan, Naphtali Rishe, Wenjia Wang, Seyed Masoud Sadjadi. arXiv preprint arXiv:2510.00376.
Wenjia Wang, Seyed Masoud Sadjadi, Naphtali Rishe. 2024 IEEE 10th Conference on Big Data Security on Cloud (BigDataSecurity), 23–34. IEEE.
Wenjia Wang, Seyed Masoud Sadjadi, Naphtali Rishe, Arpan Mahara. arXiv preprint arXiv:2405.01412.
Arpan Mahara, Naphtali D. Rishe, Wenjia Wang, Seyed Masoud Sadjadi. 2024 International Conference on Machine Learning and Applications (ICMLA), 1599–1604. IEEE.
Cristina Palacios, Jennifer Bolton, Wenjia Wang, Alayne Gatto, Lourdes Concepción, Masoud Sadjadi, Marcia Varella, Jean Hannan. International Journal of Medical Informatics, 174, 105047. Elsevier.
Wenjia Wang, Seyed Masoud Sadjadi, Naphtali Rishe. 2022 IEEE International Conference on Parallel and Distributed Processing with Applications, Big Data and Cloud Computing, Sustainable Computing and Communications, Social Computing and Networking, 262–269. IEEE.
Jennifer Bolton, Cristina Palacios, Wenjia Wang, Lourdes Concepción, Masoud Sadjadi, Marcia Varella, Jean Hannan. Current Developments in Nutrition, 6, 823. Elsevier.