01
Reproducibility
Can an AI-assisted process produce the same inspectable artifact under fixed conditions?
Research
Applied AI research with operational questions at the center.
I study the point where model output enters a real workflow: whether a process can be reproduced, whether provenance signals are trustworthy enough to act on, and how system design should change when uncertainty is material.
02
Reliability
How should a workflow behave when multimodal signals are noisy, conflicting, or incomplete?
03
Bounded action
When should an AI-assisted system proceed, refuse, or hand the case to a person?
Current publication record
The question, the current status, and why it matters
Accepted for presentation
GAISS 2026
October 28–30, 2026
Deterministic Replay Verification for AI-Generated Code Remediation: An Offline Authorization Framework
- Question
- Can repeated offline execution create stable authorization evidence before an AI-generated remediation is merged or released?
- Work
- The evaluation compares four remediation conditions across 500 synthetic Python and JavaScript examples plus a 100-example public-pattern sanity check, using five-run replay consistency, divergence, proxy rule closure, minimal-test status, and authorization outcomes.
- Product implication
- For a product team, repeatability becomes part of the release gate rather than an after-the-fact debugging concern.
Under review
IEEE Access
Submitted July 2026
PCS-R: A Reliability-Aware Hybrid Framework for Image-Text Provenance Verification and Human Review Routing in Generative AI Workflows
- Question
- Which provenance cases can move automatically, and which should stop because the available image-text signals are not reliable enough?
- Work
- PCS-R combines provenance verification with reliability-aware confidence and an explicit route for ambiguous or conflicting cases that require human review.
- Product implication
- Confidence should change workflow behavior. It should not merely become another number on a dashboard.
Under review
Journal of Systems Architecture
Current manuscript status
A Causal-Integrity Control Plane for Heterogeneous Inference Runtimes
- Question
- How can heterogeneous inference runtimes preserve causal integrity when execution crosses different models, runtimes, and control points?
- Work
- The manuscript is listed here only at its current review stage. No acceptance or publication claim is made before an editorial decision exists.
- Product implication
- The product concern is architectural: preserve enough structure around cross-runtime execution that downstream behavior can still be reasoned about and controlled.
Conference & professional service
GAISS 2026 · Austin
The accepted replay-verification paper is scheduled for presentation at the IEEE Conference on Generative AI for Secure Systems, October 28–30, 2026, at The University of Texas at Austin.
Verification
Public committee listing ↗Active research program
Agentic AI cybersecurity evaluation
This evaluation-oriented work is still in development. It examines how autonomous or semi-autonomous cybersecurity agents behave when the available information is incomplete, stale, conflicting, or adversarial, and when a privileged action should be refused or escalated.
No experiment result or performance claim is presented before the study is executed and the corresponding record exists.