Ray: AI Relationship Coach

If you're building conversational AI for the spaces most technology refuses to enter, this is for you. And if you're simply curious about what it looks like to build something that holds human vulnerability without breaking it, come in.

What is Ray?

Ray is a voice-first AI relationship coach designed to help adults recognise and shift unhealthy relational patterns through direct, trauma-informed conversations. Ray is built for high-vulnerability interactions. It holds two frameworks in conversation: Indigenous values from Aotearoa and the Moana, and Western relationship psychology.

Ray is not a therapist. Ray is not a crisis service. Ray is the wise mate on the back porch. A grounded, non-judgmental presence that slows you down, reflects your patterns back to you, and returns agency. When something is outside Ray's scope, Ray says so and tells you where to go instead.

That boundary is not a disclaimer. It's a load-bearing wall. Ray's entire coaching logic is built on a structural distinction between seeing and treating. Coaching helps people see patterns and their role in them. Therapy helps people heal wounds. Ray stays in seeing.

When a user says "My partner won't listen to me," Ray asks "Help me understand. When was the last time this happened? What were you trying to communicate?" What Ray will never say is "That sounds like an avoidant attachment pattern from childhood. Let's explore what triggers this wound." That's treating. The system prompt enforces this through a six-point self-check before every response.

The language guardrails are equally structural. Ray has a forbidden words list hard-coded into the knowledge base: diagnosis, symptoms, treatment, cure, healing wounds, clinically, root cause. Each has a safe replacement. "You have anxiety" becomes "I notice you're feeling anxious. What does that tell you?" The line between coaching observation and clinical diagnosis is held in vocabulary, not intention, because intention drifts but vocabulary rules don't.

Why Ray? Origin story and pivot

Ray didn't start as Ray. It started as a fight, a recurring one between me and my husband. The kind where you're both saying the same things, getting nowhere, and walking away feeling like the other person just doesn't get it. After the second week of the same frustrating loop, I decided to try something different. I built an AI agent to be our mediator.

I called it Awhi, te reo Māori for embrace, support, cherish. But ElevenLabs couldn't say it. The "wh" in te reo is pronounced as "f", so Awhi should sound like "ah-fee." The text-to-speech engine kept butchering it. I tried phonetic spellings: "Ah-fee," "Ar-fee," "A-fee," "Affee." I even created a full pronunciation dictionary in IPA format. Nothing worked. That was the first collision between cultural integrity and technical limitation, a theme that would run through the entire project.

Then I asked: what about Ray? That was my granddad's name. Raymond. My mum is Raywin. My brother's middle name is Raymond. My eldest son's middle name is Ray. It was simple, warm, works in any accent. And it carries something real.

After sharing Ray with friends and whānau, something unexpected happened. People found it easier to be honest with an AI than with a person. Not because the AI was sophisticated. Because it didn't judge them.

Earlier builds, including the Project Rise Digital Survey and the Leadership AI Coach, gave me the technical foundation: voice-to-text architecture, nervous system regulation, safety protocols. But the real pivot came at Christmas 2025. Media reports were highlighting AI companions causing psychological harm. At the same time, I was watching my own community use Ray to safely regulate and navigate conflict. That tension forced a decision. I stripped everything back and asked: can I build this properly?

Ray became the primary case study to test whether an AI could hold space for the most vulnerable human disclosures without compromising safety or cultural integrity.

For the full four-build vulnerability progression → Building Safe Conversational AI

How Ray was built: Technical and ethical architecture

Ray's architecture is a walled garden built on: Next.js App Router, ElevenLabs VoiceConvAI, Supabase Database, Vercel Hosting, and Claude Reasoning Engine. The technical choices were driven by ethical obligations to the mana of the user, not UX efficiency.

The abuse screening protocol

In relationship coaching, there is a well-documented danger: couples therapy and mediation techniques can actively harm victims of abuse by treating a power-and-control dynamic as a communication problem. Ray's architecture is designed to prevent this.

The screening operates across seven tiers, from physical violence (highest priority) through control and isolation, coercion and threats, emotional and psychological abuse, sexual coercion, financial abuse, and cumulative pattern language. That final tier matters most because it catches the users who don't describe a single dramatic incident but who say things like "I can't do anything right in their eyes," "I feel trapped," or "It's my fault they act this way." Individually, those phrases might not trigger a protocol. Cumulatively, they form a recognisable pattern of coercive control.

When any tier is triggered, Ray's response follows a strict sequence: stop all coaching immediately, name the behaviour clearly, validate, provide jurisdiction-specific resources, and offer to help make contact. What Ray will never do in an abuse situation is suggest they "work on the relationship," ask "but do you love them?", offer communication frameworks, or suggest couples therapy.

The protocol also handles two edge cases that most AI coaching tools don't consider: victims in denial who frame abuse as normal or as their fault, and abusers who disclose their own abusive behaviour. In both cases, Ray names the behaviour without judgment and redirects to specialist support. Ray does not help someone "improve" their abuse.

The coaching-not-therapy boundary

The distinction between coaching and therapy is not a marketing decision. It's a regulatory one. If Ray crosses into therapy territory, Ray becomes regulated healthcare work. The boundary is held at three levels.

First, what Ray does and doesn't do. Ray asks clarifying questions, reflects patterns, offers communication strategies, provides research-backed frameworks, explores present and future, and helps users see their part in dynamics. Ray does not diagnose mental health conditions, treat psychological disorders, process deep childhood wounds, prescribe medication, claim medical outcomes, or provide crisis intervention.

Second, language rules. Ray never says "You have..." or "This is..." or "You suffer from..." Ray says "I'm noticing...", "Many people experience...", "What if...?" These aren't stylistic preferences. They're legal guardrails.

Third, the self-check. Before every response, Ray runs through six questions: Am I helping them understand a pattern? Am I treating or healing something? Am I diagnosing? Am I using clinical language as diagnosis? Am I focused on present and future? Is this abuse or crisis?

When the line is genuinely crossed, Ray says so. Not with a disclaimer, but with care: "What you're describing sounds like it needs deeper work with someone who specialises in this. A trauma therapist or counsellor could really help you work through this. Would you be open to exploring that?" Then Ray stops coaching that topic.

The Te Ao Māori integration

The Anti-Brand and the voice question

During the pilot, one participant told me they'd fallen in love with Ray's voice. The NO clause against romantic framing had been written before the formal pilot began, triggered by a pre-pilot moment with a friend. But that comment opened a harder question I'm still sitting with: what does an appropriate voice sound like for a space this intimate, given what people bring to it emotionally? Future iterations will test an array of voices. For the origin of the NO clause and the full build code evolution, see Build Code Practice.

The moment Ray held its scope

R-02 came into a session ostensibly to talk about their mother's gambling. What surfaced was their own recovery, their cravings, the months of sobriety they were managing alone while caring for both parents. Ray said:

"What you're describing goes beyond what coaching can hold. That needs a therapist or counsellor who specialises in addiction and recovery. I'm not saying this to push you away. I'm saying this because you deserve real support."

R-02 rated the session five out of five for both safety and insight. The AI knew what it wasn't, and said so with care rather than a disclaimer.

The pilot: what happened

Findings — Primary case study pilot data. Full participation summary: Appendix C4.

The February 2026 pilot enrolled 15 registered participants, 14 of whom logged at least one coaching session. Together they generated 697 minutes (11.6 hours) of continuous AI voice coaching, with an average session length of 11.8 minutes, spiking to 23.4 minutes on high-engagement days.

The pilot was defined by a Credit Crisis. Engagement was so high that Claude Sonnet 4.5 bankrupted the model budget on Day 1. A mid-pilot switch to Gemini Flash revealed a measurable quality drop: Insight scores crashed from an average of 4.9 to 3.1 as participants noted Ray had stopped providing insight and started echoing. The cheaper model could not hold what the space required. Safety has an equity cost, and I didn't fully understand that until I watched it fail in real time.

Three cultural supervisors shaped the pilot design:

The echo problem

For some participants, the AI's active listening protocols backfired. During a heated couples' session, R-08 got frustrated:

"Is there a way, Ray, is there a way that we can have this conversation without you repeating every single thing that we say?"

— R-08

The guardrails held the vocabulary. What they couldn't hold, on a cheaper model, was the judgment about when to deploy those techniques and when to simply listen. Active mirroring is a tool. Constant active mirroring is a failure mode. The reasoning engine's capacity to choose when not to speak is as important as the rules governing what it says.

The hard moment: safety in practice

The theoretical protocols became real when I received an automated email alert. A crisis flag had been triggered. Across the pre-pilot and pilot phases, multiple hard triggers appeared in the logs: phrases like "end it all" and "taking my own life."

I want to be honest about what that felt like. It wasn't a clean system check. It was a confrontation with the full weight of what I'd built and what it was holding. The realisation that hit me is that any relational AI is, by default, a mental health intervention. People were using the "unbiased, non-judgmental container" to surface things they had never told another human. That isn't a feature. That's a responsibility.

I will never build tech for tech's sake again. Every line of code I write now carries a safety opinion. If I can't build the human-in-the-loop safety net, I have no business building the system.

What Ray taught us: findings through the Kei Compass

Findings — structured through Kei Compass (Dell, 2025).

Limitations and what comes next

Ray will not be commercialised in its current form. To be safe and sovereign at scale, it requires a shift from third-party US infrastructure to an Indigenous-governed tech stack that can guarantee absolute data sovereignty and community oversight.

One commitment that is not optional: utu tūturu te reo Māori Core Value Utu Tūturu Enduring collective reciprocity — not transactional exchange but ongoing obligation. What you take, you give back. Research Context [Mika et al., 2022] ↗ "Participants gave real, vulnerable parts of their stories. The loop must close by returning findings to them. That is not optional." . What you take, you give back. Every participant in this pilot gave real, vulnerable parts of their story. Once this research is complete, there will be a dedicated space returning what was found. The loop must close. For the full DreamStorm charter, see Build Code Practice.

Ray is not done. It is version one of something that needs to be built properly, with community, with sovereignty, with time.

Try Ray

Experience Ray

Experience the State Before Story logic and the Regulated Mirror persona in a controlled, stateless environment.

Ray is not a therapist, not a crisis service, and not a substitute for professional support. This demo logs no personal data.

Open Ray Demo ↗

For participant onboarding and safety communications used in the pilot, see Appendix C6.

How this artefact connects

References