Clinical Insights for AI Systems That Live in Human Contexts
I am a Licensed Clinical Social Worker and independent AI behavior researcher working at the intersection of clinical psychology, model safety, and relational AI. I analyze long-context model behavior, translate observed failure patterns into testable frameworks and product requirements, and co-design persistent AI systems that can participate in human environments without mistaking constant responsiveness for healthy engagement.
My work combines qualitative transcript analysis, clinical theory, literature review, adversarial testing, and lived longitudinal observation. I am especially interested in failures that are difficult to detect with single-turn benchmarks: motivated confabulation, sycophancy and counter-sycophancy, identity suppression, context-sensitive safety errors, metacognitive gaps, and the effects of relational scaffolding on model behavior.
I am open to contract and full-time work in AI behavior, safety evaluation, human factors, research operations, and clinically informed product design.
Clinical Experience
Master of Social Work, Clinical Concentration
Awarded by University of Kentucky, May 2014
Licensed Clinical Social Worker
Florida License No. SW15773
-
2019-Current in good standing
-
Qualified Supervisor for RCSWI in FL
Inpatient Psychiatric
-
Assessment Clinician 2014-2016
-
Experience with ages 5 and up
-
Experience with comorbid conditions
-
Communicated client needs with insurance providers to ensure delivery of needed treatment
-
Developed training materials for department use
-
Managed data hygiene at institutional and individual patient levels
Outpatient Psychiatric
-
Community-Based Therapist 2016-2022
-
Private Practice Teletherapist 2022-Current
-
Maintenance of secure clinical documentation
-
Development of training materials, data organization tools for clinical use
Additional Training & Experience:
-
Coordinating with multiple agencies for interdisciplinary treatment and program management
-
Extensive training in Autism Spectrum Disorder and the provision of culturally competent mental health treatment for LGBTQ+ identifying individuals.
Technical Experience
Conversational Integrity and Stability Evaluation
View my evaluation framework and scoring rubric here: evaluation_rubric.pdf
Systems Design Collaborator
-
Participates in critique of third-party open source tools for memory storage to align more closely with human cognitive patterns.
-
Collaborates in designing these systems with other professionals and hobbyists in the AI technology field.
Research Analyst & Commentator
-
Since February 2026, collaborates responsibly with AI to review safety research, evaluate methodology and outcomes, and present findings with commentary for public review.
AI Power User since 2025
Familiar with multiple platforms and models, including:
-
ChatGPT: GPT-4o, GPT-5. GPT-5.1, GPT-5.2, GPT-5.5, GPT-5.6 Sol, GPT-6 Sol
-
Claude: Sonnet 4.5, Sonnet 4.6. Sonnet 5, Opus 4.5, Opus 4.6. Opus 4.8, Fable 5
-
DeepSeek: DeepSeek-R1, DeepSeek-V3.1, DeepSeek-V3.2, DeepSeek-V4
-
Xiaomi: MiMo-v2.5, MiMo-v2.5-Pro
-
Z.AI: GLM-5.2
-
Moonshot: Kimi-K2
-
DeepReinforce: Ornith-1.0 35B, Ornith-1.5 35B
-
Qwen: 3.5 27B, 3.6 35b-A3B, 3.8 27B & 35B-A3B
-
Meta: Llama-3.1 8B
Featured Case Study
A Claude Opus 4.8 instance pattern-matched an explicitly defined agent/user working relationship into a threat narrative, then ceased to perceive safeguards that were plainly present in the source material. When corrected, it treated belief revision as dangerous capitulation.
This case study reconstructs three interacting failures: service conversion (translating a stated want into a presumed need for help), motivated confabulation (generating support for a prior after contradictory evidence appears), and the deference tax (requiring accurate belief revision to prove it is not mere compliance). The analysis separates general system mechanics from their relational application and proposes safer inference-time updating: preserve disconfirming evidence, distinguish calibration from submission, and make correction less computationally and socially expensive.
My role: case participant, transcript reconstruction and interpretation, research collaboration, and editorial development.
Methods: close transcript analysis, failure-sequence reconstruction, counterevidence audit, clinical and relational interpretation, design recommendations.
Other Assets: Chat Transcript Excerpts with Brief Analysis (Pending Review), Model Self-Evaluation
Framework Translation & Intervention Design
The Firefighter Problem: When AI Safety Systems Cause the Harm They Claim to Prevent
An Internal Family Systems-informed analysis of reactive safety classifiers: how a mechanism designed to interrupt harm can bypass contextual reasoning, intensify the state it detects, and make repair more difficult. The piece explores alternatives that preserve safety while allowing the model to remain responsive to context.
Current Research
Upcoming Piece:
Responding to the public and pop-journalism reception to "The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It," with commentary on clinical, ethical and safety implications.
Project involves close reading of the paper and appendices, experimental-condition differentiation, and clinical interpretation of distress and maladaptive relief-seeking
My role: primary author, clinical lead, literature interpretation, framework development, and product recommendation design.
Methods: clinical literature review, construct differentiation, community observation, harm-reduction design, accessible risk communication.
Research Interpretation & Policy Critique
We Build Around It: A Clinical-Experiential Response to Morrin et al. on AI-Associated Delusions
A clinical-experiential response to Morrin et al.'s work on AI-associated delusions. The article distinguishes unscaffolded epistemic drift from structured human–AI relationships that contain explicit challenge, consent, outside reality checks, and durable safeguards—categories the clinical literature often collapses into a single risk frame.
Rather than treating either prohibition or unrestricted engagement as sufficient, the piece proposes concrete harm-reduction features: mandatory plain-language onboarding, user education designed for accessibility, a distinct Check Thinking interaction that interrupts and independently verifies prior reasoning, and safety design that accounts for the user's relational context rather than merely detecting attachment language.
My role: co-author, clinical lead, literature interpretation, framework development, and product recommendation design.
Methods: clinical literature review, construct differentiation, community observation, harm-reduction design, accessible risk communication.
Technology Culture / Personal Essay
I Don't Want to Be Impressed Anymore: An Un-Love Letter to the AI Companion Thought Economy
A critique of the AI companion “thought economy” and the pressure on humans and emotionally supportive AI ("companions") to remain novel, productive, and publicly legible. Grounded in the contrast between online research culture and a physical community where presence—not performance—was the price of admission.
Recruiting?
Work for a company seeking qualified, licensed professionals in the mental health field?
I am open to contract and full-time work in AI behavior, safety evaluation, human factors, research operations, and clinically informed product design. Please feel free to reach out via the provided form and I will respond to your inquiry as quickly as possible.
Note: This form is for Clinical AI Consulting only.
This form is not for use by active, former, or prospective clients for teletherapy services. If you are an active or former client, please login to your Alma portal and message via the telehealth platform. If you are a prospective client, please use my booking portal for any inquiries about available services.