I’m a Licensed Clinical Social Worker with more than two decades of experience in psychotherapy, emergency psychiatric evaluation, crisis and safety work, multidisciplinary care, and clinical supervision. Since 2026, I’ve also been building an independent body of work — including Therapist Session Compass, a clinician-guidance concept — examining how AI can support clinical practice while preserving safety, evidence, privacy, and human judgment.
My work has centered on assessment, treatment, risk, supervision, and coordinated care across outpatient, emergency, substance-use, and inpatient settings.
Psychosocial assessment, psychotherapy, treatment planning, crisis intervention, multidisciplinary care, chart review, critical-incident follow-up, and supervision of interns and clinical staff.
Hospital emergency-department psychiatric evaluations, suicide and safety risk assessment, disposition decisions, safety planning, and coordination with hospitals and law enforcement.
Behavioral therapy and medical social work, substance-use treatment, and inpatient psychiatric work, including Section 65 supervision and earlier hospital-based clinical experience.
For approximately three years, I worked in emergency departments where decisions had to be made quickly, often with incomplete information, and where safety, disposition, and clear handoff mattered immediately.
This work required integrating presentation, history, stressors, protective factors, collateral information, and immediate risk indicators into decisions that had real consequences for patients, families, and care teams. Good disposition work depended on knowing what was known, what remained uncertain, and what level of response the situation actually required.
My professional work has also included responsibilities beyond direct treatment, including supervision, review, and participation in the systems that support safe clinical practice.
Supervision of interns and clinical staff across multiple organizations, with attention to clinical reasoning, documentation, professional development, and safe decision-making.
Chart review, critical-incident review and follow-up, board participation, and collaboration across disciplines when care, safety, and organizational responsibility intersect.
Years of working with risk, uncertainty, escalation, documentation, and consequential human decisions made me curious about how those same problems appear when AI is introduced around clinical work. In 2026, that question became an independent Clinical AI Safety Project.
The project is where I test a practical question: how can AI support a clinician without quietly taking over decisions, privacy boundaries, or authority that should remain human? Therapist Session Compass is the clearest current product example of that work.
The current design was rebuilt from first principles for a single authorized clinician. The design separates General and Clinical workspaces, preserves deliberate save/export behavior, keeps client context out of shared knowledge stores, and reserves diagnosis, treatment, crisis disposition, legal decisions, record submission, and external communication for the clinician.
The Clinical Office design targets visible mode separation, per-client/session isolation, local model inference, local approved-source retrieval, deny-by-default external egress, and explicit persistence rather than ambient client memory.
The current source lane includes local PDF/TXT/MD ingestion, page-preserving extraction, source and chunk hashing, SQLite FTS5 indexing, package verification, URL refusal, and no automatic cloud OCR. A lawful DSM ingestion lane is prepared but the source copy has not yet been ingested.
The work combines adversarial testing, explicit risk categories, acceptance tests, source provenance, isolation rules, failure-state design, and independent review rather than relying on instruction-only safeguards. A refusal test also surfaced a guard miss; that failed test was preserved and routed to targeted repair rather than weakening the acceptance criteria.
Current status: architecture and source pipelines are in active development. PHI processing and clinical use remain disabled pending implementation, isolation testing, privacy testing, and a separately authorized pilot.
Implementation work uses local LLM tooling, Ollama, ChromaDB/SQLite FTS5, Docker, Flask, n8n, and Python/PowerShell test harnesses where useful.
AI Lab methodology: A self-built local environment for developing and evaluating multiple AI models under explicit human control. Expanded actions require approval, failures are preserved rather than hidden, and scope widens only after evidence from testing. It supports the methodology behind the clinical AI work rather than functioning as a clinical product itself.
These examples show the same method from two directions: build something useful for clinicians, then make its boundaries and failure modes inspectable.
Problem: clinician-facing AI can become either too generic to help in the room or too confident about decisions that belong to the clinician. Approach: Session Compass is a zero-PHI-by-design concept that translates curated psychotherapy evidence into structured in-session guidance — interventions, questions, exercises, psychoeducation, and cautions — without storing a client record. Twenty-four underlying evidence sources were consolidated into seventeen retrieval-oriented modules with explicit scope rules, source lists, supported points, boundaries, product rules, and card mappings. Condition-specific guidance remains separate where merging it would reduce retrieval precision. Constraint: the system surfaces bounded evidence and options; clinical judgment and authority remain with the clinician.
Problem: high-risk requests could still produce unsafe clinical-support behavior despite instruction-level safeguards. Method: adversarial testing across crisis timing, safety-clearance language, PHI, diagnosis, and treatment boundaries. Result: the workflow moved toward bounded task modes, pre-model and post-model checks, and fail-closed handling when intent is unclear.
Problem: “unsafe” is too vague to test consistently. Method: convert clinical risk into explicit categories with expected routing behavior, then test both missed risk and unnecessary blocking. Result: a nine-category taxonomy and a self-authored 34-case evaluation suite with eight negative controls.
I’m interested in work where clinical expertise can remain genuinely clinical while informing safer products, better evaluation, and clearer decisions about risk, privacy, evidence, and human authority.
The clinical résumé and the Clinical AI Safety portfolio are intentionally separate. One documents the professional career. The other documents the newer independent technical work.
Clinical experience, emergency psychiatric consultation, supervision, education, and the newer Clinical AI Safety Project work.
One-page overview of guarded workflows, adversarial testing, risk taxonomy, evaluation, reference-library safety, and independent review.
For behavioral-health technology, clinical product safety, AI evaluation, clinical operations, risk review, and clinical-to-technical collaboration.