SDSU AI4Business Lab
Exploring the gap between trust in AI and its technical trustworthiness
My work spanned disaster-response information systems, AI-security literature, adversarial-attack experiments, and agent controls. In my second year, I founded and led a research team investigating the relationship between human trust in AI and technical reliability and security.
My contributions: research synthesis and framework development · team leadership · experimental tooling and reproduction · agent prototype design and implementation
Connecting human trust with technical evidence
I developed a three-layer framework connecting psychological trust, technical safety and robustness metrics, and security threats. Alongside literature review and an attack-method taxonomy, this organized our investigation of how people’s confidence in AI relates to its technical properties. I contributed to the team’s manuscript, with substantial writing and discussion contributions from other members; it was shared internally within the lab.
As the work progressed, my interests expanded from model robustness toward human overtrust and system controllability.
Organizing the investigation
I founded and led the AI-security team, which grew from three to four members across two semesters. I designed a paper-review process with a five-axis classification scheme, a shared glossary, structured reading notes, and an assignee–second-reviewer workflow.
Early work showed that we needed a stronger shared foundation before applying the classification consistently. I shifted the approach toward foundational papers and reading groups. The planned mapping dataset was not completed; the methodology and the revised research process were the work delivered.
Investigating model robustness hands-on
I reproduced FGSM and PGD adversarial attacks on Fashion-MNIST using TensorFlow and the Adversarial Robustness Toolbox. My experiment-note template documented objectives, hypotheses, success criteria, environment details, attack settings, results, and visualizations of clean images, adversarial images, and perturbations.
At perturbation budget ε = 0.05:
- FGSM: accuracy fell from 88.05% to 29.55%.
- PGD: accuracy fell from 88.50% to 19.35%.
The experiments used the same architecture, training recipe, and four-value perturbation grid, but separately trained model instances. They reproduced known attack behavior rather than establishing a new result.
I also built a Docker environment with CPU/GPU targets, Compose profiles, and pinned dependencies to support repeatable experiments. We presented progress at the lab’s end-of-semester sessions.
Exploring controllability through an agent prototype
I wrote a LangGraph Secure Sentinel specification for the team and implemented S3 Sentinel v1 as a worked example. The S3-themed prototype explored role-based tool access, a gatekeeper node, human-approval breakpoints, and checkpointed state. Roles came from application configuration rather than claims in a user prompt.
The implementation also added response sanitization and audit tags. This was an exploration of agent controls, not a production-security validation; the later AWS implementation is covered in the separate S3 Sentinel project.
Earlier work: disaster-response information
In my first year, I reviewed literature on AI-assisted disaster information, including BERT models and knowledge graphs for accuracy and context-awareness.
Publication: Listed coauthor, “Agentic AI for Crisis Informatics: A Multi-Agent Framework for Scalable and Reliable Disaster Communication,” AMCIS 2026 — selected among the conference’s Top 25% Papers. My involvement was in the earlier literature-review phase; the published framework evolved afterward.