Psychology · AI Governance · Research
I turn AI regulation into systems teams can run.
Incident frameworks, policy gap analyses, evaluation plans with release gates, and four live tools, grounded in NIST AI RMF, ISO/IEC 42001, OSFI E-23, the EU AI Act, and Canada's AIA. Built on an AI Risk Governance co-op at Rogers and a psychology research background that takes the human side of AI risk seriously.
Available for Fall 2026 and Summer 2027 co-ops.
AI Governance
& Risk
I make AI governance less abstract — turning frameworks like NIST AI RMF, ISO 42001, and the EU AI Act into structured documentation, incident frameworks, and process guidelines real teams can use.
Research
& Analysis
Across multiple psychology labs, I've developed the habit of turning messy information into something clear, organized, and actually useful — whether that's data, literature, or a dashboard.
Human-Centered
Systems
I care about who gets left behind. My work asks not just "does this function?" but "who does this help, and what would make this easier to trust?"
Most problems aren't technical — they're about translation. Between policy and practice. Between data and decision. Between systems and the people inside them.
My approach
to messy
problems.
I'm usually the person organizing documents, clarifying expectations, tracking patterns, and translating complexity into something other people can actually use. Quietly, carefully, and with a lot of notes.
Before anything else — what do I actually have? What's the real scope? I'd rather spend time understanding the problem than solving the wrong one.
Some problems need diverse input. Others need one person to just go deep. I've learned to tell the difference.
I create frameworks before I create outputs. Documentation, templates, synthesis — the scaffolding that makes complex work navigable.
A solution that can't be explained clearly hasn't been fully solved. I iterate until the clarity matches the quality of the work.
the following is a dramatisation.
caffeine and chaos were involved.
Designed an AI Incident Reporting Template and reference architecture aligned to NIST AI RMF, ISO 42001, OSFI Guideline E-23, and the EU AI Act. Conducted a comparative AI policy gap analysis benchmarking the corporate AI policy against Government of Canada AI frameworks; proposals being considered for the next policy version. Developed a pre/post-deployment Responsible AI evaluation plan for an internal AI tool, including PASS/FLAG release gates and monitoring KPIs. Authored a plain-language guide translating seven Responsible AI principles into accessible workplace guidance for non-technical employees.
Six-person student consulting team engaged to evaluate AI chatbot integration for a Canadian healthcare staffing platform. Authored the Workforce Analysis section of the final client report, examining user values and identifying four AI-driven workforce experience opportunities. Contributed to industry research, competitive analysis across five direct competitors, and solution evaluation across three integration approaches (AWS Lex, Stream Chat with LLM integration, Yellow.ai).
Conduct literature reviews and contribute to study design discussions on social perception research. Support data collection and entry under graduate student supervision. Maintain documentation aligned with TCPS 2 research ethics standards.
Recruited participants and administered study protocols handling confidential research data. Contributed to study design discussions, prepared experimental materials, and completed qualitative data coding. Supported analysis of participant data on intergroup relations. Operated under TCPS 2 research ethics standards throughout participant engagement.
Co-led a creative initiative built around peer interviews and shared work-experience storytelling, which grew to a 50+ person cross-functional team across multiple time zones. Screened 500+ applicants, led onboarding, and built operational structure for an all-volunteer team. Worked with the creative team on design reviews — providing structured critique on app interfaces and logos. The initiative was paused before public launch at the end of 2024; the operational foundation, the practice of interviewing strangers, and the design-critique muscle remain some of the most valuable things I took from undergrad.
Serve on the MindPad Editorial Board reviewing undergraduate and graduate research submissions against APA standards and ethical research guidelines. Provide developmental feedback on methodology, clarity, and writing to support author revisions. Also serve as Student Representative for the CPA Section for Students, representing undergraduates nationally and directing peers to academic and professional resources.
Co-designed mental health awareness initiatives, created outreach materials, and supported help-seeking through tabling, residence pop-ups, and workshops tailored to diverse student groups.
Do Large Language Models Know When They Are Wrong?
A completed study on LLM confidence calibration: when a model says it's "highly confident," is it actually more likely to be right? It measures the gap between how certain a model sounds and how accurate it is — the overconfidence bias psychology has studied in people for decades, now pointed at AI — across a prompt × domain × model design spanning psychology, health, and finance. My research partner and I put 89 expert-level questions to Claude Opus 4.8 and GPT-5.6 under three prompts: 534 responses in all.
Both models were right about nine times out of ten, then missed their own accuracy in opposite directions: Claude underconfident, GPT overconfident. Telling them to be humble lowered stated confidence significantly (p = .004) and made calibration worse for both, because lowering confidence is a direction, not a correction. A model that is confidently wrong is a problem; a model whose confidence has nothing to do with whether it's actually right is a safety problem — and no single prompt fixes it.
× domain (psychology · health · finance)
× model (Claude Opus 4.8 · GPT-5.6)
- Claude: 89.9% right, 86.8% claimed — underconfident.
- GPT: 92.1% right, 96.2% claimed — overconfident.
- A humility prompt helped one and hurt the other.
Is It Smart, or Does It Just Sound Like Us?
A standalone interactive essay on AI through a psychology lens — five questions I investigated, built around my own Turing-test study in which an AI told to "sound human" was judged more human than me (150 vs 101). Visitors can run a live experiment: be scored on how human they sound, or try to spot the human better than the AI judge can. I designed the concept, ran the study, and wrote every argument and section; the Next.js site was built using AI-assisted development.
AI Governance Audit Framework
A 30-question governance audit across 6 domains — including a dedicated OSFI E-23 domain. Estimates Canada AIA Impact Level (1–4), generates a 16-KPI governance blueprint across pre/post/continuous phases, and maps every finding to its regulatory clause. I designed the full regulatory architecture, scoring logic, and report structure; the governance knowledge driving it came from my AI governance work and independent research into Canadian regulatory frameworks. The Python and Streamlit implementation was built using AI-assisted development.
Universal Analytics Engine
A dataset-agnostic BI platform with a built-in Model Risk Governance layer designed around OSFI E-23 and Canada's AIA. I designed the governance architecture: PII detection logic, AIA Impact Level scoring, proxy bias flagging, data integrity scorecard, and downloadable compliance report structure. The Python and Streamlit implementation was built using AI-assisted development; the differentiated work is the regulatory architecture above.
ESG Corporate Risk Scorecard
A sector-adjusted scoring tool evaluating organizations across 20 ESG criteria using SASB materiality weights. Financial Services prioritizes Governance; Energy prioritizes Environmental. Includes TCFD alignment, GRI Standards coverage, and sector-specific material risk cards. I designed the sector materiality mapping and scoring logic; the Python and Streamlit implementation was built using AI-assisted development.
Currently
Exploring
Unfinished work is still real work. These are the ideas I'm actively thinking through.
AI Incident Management Frameworks
Exploring how organizations can structure AI incident response — what gets logged, who owns it, and how governance intersects with technical failure modes.
AI in Healthcare Platforms
Continued thinking on where AI meaningfully improves complex workforce platforms versus where it introduces new risks — drawing on research from a UWaterloo consulting program.
Unity Companion AI Prototype
A third-person gameplay prototype in Unity (player movement, camera control, enemy AI, companion behaviour). Unlike my governance tools, I wrote the C# here myself, working through it script by script with AI as a tutor, and deployed a playable build. Currently paused. The project where I've learned the most about reading and debugging my own code.
Why this
combination
matters.
Most people who work in AI governance come from law or engineering. I come from psychology — the science of how people actually think, decide, and behave. That's not a gap. That's the point.
- Psychology + AI Governance: I understand both the technical frameworks (NIST, ISO 42001, EU AI Act) and the human factors those frameworks exist to protect.
- Research rigor with practical output: I've worked in academic labs and corporate governance — I can read a methods section and write a governance template.
- I ask the uncomfortable questions: Not just "does this work?" but "who does this work for?" and "what happens when it doesn't?"
- Translation as a skill: I move between technical complexity and human legibility — a rare combination that matters in any risk or governance role.
- LLM Evaluation
- Confidence Calibration
- Red-Team Thinking
- Model Behavior Analysis
- PASS/FLAG Release Gates
- Incident Frameworks
- NIST AI RMF
- ISO/IEC 42001
- EU AI Act
- OSFI Guideline E-23
- Canada AIA
- AI Risk Management
- Responsible AI Auditing
- Policy Documentation
- Literature Reviews
- Qualitative Coding
- Research Methods Foundation
- Data Interpretation
- Academic Writing
- Synthesis & Reporting
- Editorial & Peer Review
- Python · Pandas (analysis & prototyping)
- Streamlit · Plotly
- Microsoft Azure (AI Fundamentals certified)
- Microsoft 365 / SharePoint
- Notion · Zotero · LaTeX
- Figma · Miro
- Stakeholder Communication
- Project Coordination
- Cross-functional Collaboration
- Design & Framework Critique
- Mental Health Literacy
- Equity & Inclusion Work
Let's
work
together.
I'm looking for Fall 2026 and Summer 2027 co-ops in AI governance, responsible AI, technology and model risk, AI evaluation, and AI product. If you're hiring in that space, email me and I'll reply within a day.