AI Safety Researcher · Cognitive Systems Designer

Pooja
Rangarajan

Applying how humans regulate distress to how AI systems regulate behavior.

I work at the intersection of psychological therapeutic frameworks and LLM safety evaluation — building the case that AI safety is a behavior problem, not a content-filtering problem.

My research introduces computational psychopathology as a new area: using clinical models of human distress regulation to architect how AI systems modulate responses under adversarial or high-stakes conditions. Published on arXiv, filed as a provisional patent, tested through frontline crisis work and independently deployed tools.

arXiv paperpublished · 2025
provisional patentfiled · active
crisis intervention200+ hours
tools deployed2 live
arXiv Paper ↗ LinkedIn ↗
scroll
Research & Credentials

Work that exists
in the world.

Published Paper · arXiv 2025
Mitigating Harmful Erraticism in LLMs Through DBT-Based De-escalation Strategies
Proposes applying Dialectical Behavior Therapy principles — mindfulness, distress tolerance, emotion regulation, interpersonal effectiveness — as a regulatory framework for LLM behavior. Introduces a computational distress function and SUDS-based intervention architecture. Co-authored with MARCo Health. Coins the term computational psychopathology as a new research area.
Read the paper
Provisional Patent · Filed
Applying Psychological Therapeutic Principles to AI Safety
A provisional patent formalizing the application of therapeutic modalities to AI safety architecture — bridging clinical psychology and LLM alignment research. Grounded in the same theoretical foundation as the arXiv paper, with emphasis on practical implementation pathways.
Research Internship · MARCo Health
LLM Behavior Evaluation in Mental Health Contexts
Evaluated LLM behavior and safety in mental health AI at MARCo Health Inc., Newark NJ. Contributed to research on response modulation, user vulnerability detection, and the failure modes of therapeutic chatbots. Directly informed the paper's empirical methodology.
Crisis Intervention · 200+ Hours
Crisis Text Line & NAMI Frontline Practice
200+ hours of active crisis intervention as primary domain validation for AI safety research. Understanding what de-escalation requires in practice — and precisely where AI systems currently fail — is the empirical ground the theoretical framework is built on.
Implemented Tools

Proof of concept,
deployed.

Two tools built independently to validate the core research claim: that state-aware, regulation-first AI design produces meaningfully different outcomes than standard chatbot architecture. Each is a design artifact demonstrating nervous-system-aware routing, persona architecture, and safety boundary formalization in working code.

01
CALM OS
Emotional State Routing System
Open Live App ↗
Vanilla JSState RoutingCrisis DetectionVoice OutputDBT-InformedClient-Side Only
An emotional state-routing companion built on the research principle that regulation must precede action. Uses a keyword classifier and session-aware logic to route users through one of three state pathways (red / yellow / green), adapting phrasing, response timing, and escalation handling in real time.

Includes crisis detection with inline resource cards. The animated orb UI and task breakdown flow (regulate → release → do) operationalize the paper's core architecture in a browser-native, zero-data-collection implementation. All processing is client-side — no server, no tracking, no storage.
02
COMPASS
Clinical Support System Prompt
System Prompt DesignPersona ArchitectureMode SwitchingSafety Boundaries
A multi-modal system prompt framework demonstrating how persona architecture, behavioral mode switching, and formal safety boundaries can be specified within a single prompt document. Built around an EMH (Emergency Medical Hologram) external persona with Sherlock-style internal deductive logic.

Three operational modes operationalize the paper's state-routing principles in prompt form. A design artifact as much as a technical one: proof that safety architecture can be implemented at the specification layer.
Frameworks & Methodologies

Applied cognitive science,
documented as transferable systems.

Two frameworks developed through autoethnographic research methodology — using the researcher as primary subject to generate domain knowledge that external observation cannot produce. The result is transferable, documented systems with defined inputs, outputs, and failure modes.

Original Framework · December 2025
Synesthetic Chaining System (SCS)
A sensory-sequencing framework for task execution and executive function scaffolding — designed for cognitive profiles that process through sensation rather than verbal planning. The core insight: instead of asking "what do I do next?", the sensory environment itself provides the signal.

Built on the principle that executive function failure is often a signal-routing problem, not a motivation problem. SCS replaces checklist-based task management with linked perceptual chains, reducing cognitive overhead by embedding decision logic into the environment. Three chain types:
Chain TypeFunctionExample
NavigationLocation → next step signalBed → floor = sand → door = white boundary → bathroom = pearl / aquatic
NarrativeEnvironmental sequence → action orderWide messy space → narrower → refined → clean surface
Paired RegulationUnpleasant input → calming anchorCold → volcanic containment · Rough → rock grounding
Workbook authored by Pooja Rangarajan · December 2025 · Educational use only. Not a medical treatment or clinical intervention.
Personal Learning Framework · Applied
Visual-Spatial & Synesthetic Learning System
A learning methodology built from systematic observation of how information is processed across three cognitive channels simultaneously — visual-spatial, auditory-verbal, and kinesthetic-empathic. The framework formalizes which channel to engage for which material type, and how to route between them to maximize retention and reduce overload.

The central finding: most standard instruction is optimized for a single channel. Multi-channel learners fail not from incapacity but from channel mismatch. The system provides two techniques that activate the correct channels for the material type:
Technique 1 — Immersive Scene Learning
Best for: concepts, formulas, understanding why something works

1. Set a consistent home base scene once — same environment every session. Don't rebuild; populate with new concepts.
2. Place the concept inside the scene visually. A formula becomes a holographic display; a relationship between variables becomes two characters interacting.
3. Feel the environment — atmosphere, urgency, temperature. The kinesthetic channel locks the material in.
4. Explain it out loud in one sentence as a character inside the scene. Auditory channel reinforces the visualization.

Example: P-value → a patient's symptom on a holographic display. Low p-value = the system takes it seriously. High = probably noise.
Technique 2 — Synesthetic Color-Shape Anchoring
Best for: numbers, calculations, memorizing specific values

1. Notice the automatic association — look at a number and observe what color, shape, or texture arises. Don't force it.
2. Trust what comes — synesthetic associations are consistent and reliable across sessions.
3. Let them blend — when combining numbers, observe what the blend produces. The result typically arrives before conscious calculation completes.
4. Reinforce once verbally to engage the auditory channel alongside the synesthetic one.

Example: 6 = pink, 7 = purple. When they merge, dark pink arises — and 13 arrives with it. Calculation through perceptual blending, not arithmetic.
Quick Reference
Use Immersive Scene when learning what a concept means or why it works.
Use Color-Shape Anchoring when memorizing specific values or doing mental calculations.
Personal methodology document · Educational use only. Not a prescribed treatment or diagnostic tool.
Design Philosophy

How I think about
AI and mental health.

01
State-first, always.
Before asking what a user wants, a system should detect where they are. Dysregulated users don't need more choices — they need fewer. Every tool I build routes on state before intent.
02
Regulation before action.
DBT's core insight: you cannot engage effectively when dysregulated. AI systems that skip this — jumping to problem-solving during distress — replicate exactly the failure modes in the clinical literature.
03
Presence over advice.
The most harmful thing a mental health AI can do is over-validate or prematurely reassure. What actually helps is acknowledgment, containment, and one next step.
04
Restriction isn't safety.
Current AI safety treats harm avoidance as content filtering. My paper argues systems need response modulation capacity, not just guardrails. Safety is a behavior, not a blocklist.
05
Build for the hard states.
Most tools are designed for people who are already regulated. The design decisions that actually matter are the ones made for sensory overload, shutdown, and executive dysfunction — the states where current AI fails.
06
Credibility through specificity.
Vague claims about "responsible AI" are everywhere. What I care about is testable: this framework reduces erratic outputs, this routing logic changes outcomes, this architecture has measurable safety properties.
Contact

Let's talk about what
AI should do
for people in distress.

Open to roles in AI safety evaluation, HITL specialization, and mental health AI research. Also open to research collaboration and contractor work.

Paperarxiv.org/abs/2510.15889→ LinkedInlinkedin.com/in/pooja-rangarajan→ Emailpmr117@sph.rutgers.edu→