REAA: Emotionally Attuned
AI Therapy Assistant
A clinically safe AI system for mental health support — validated through rigorous academic research, engineered for deployment at scale.
A Critical Gap in Mental Health Care
Over one million people in England are currently on NHS mental health waiting lists, often facing delays of months or years. Meanwhile, people are turning to unregulated AI chat systems for emotional support — products with no clinical oversight, no safety mechanisms, and no accountability.
The opportunity is clear: an AI system that delivers supportive dialogue which is both clinically meaningful and demonstrably safe could address a critical gap in care. But the stakes are too high to follow the typical software startup playbook. In mental health, getting it wrong causes real harm.
REAA takes a different approach: validate safety through academic research first, then build toward deployment. This isn't slower — it's the only path that meets the standards healthcare partners will require.
What REAA Is
REAA (Real-time rendered Emotionally Attuned AI therapy Avatar) is a multimodal AI system that combines speech recognition, emotionally aware dialogue, and a digital avatar to deliver supportive conversational interactions. It is governed by a deterministic safety stack — rule-based controls that operate independently of the underlying language model.
The system is being validated through formal academic research at Aberystwyth University, where trained psychotherapists engage in structured sessions with the avatar. This research provides the evidence base that healthcare partners need before deploying AI in clinical pathways.
The long-term target is a deployable tool for NHS digital mental health pathways, private clinics, and academic institutions in the UK and beyond.
Built for Safety, Designed for Scale
REAA's architecture is built on a modular Python 3.12 stack with local inference, deterministic safety controls, and a renderer-neutral avatar system. The same core that runs the research instrument scales to cloud deployment for clinical use.
Multimodal Pipeline
Real-time speech-to-speech processing with voice tone analysis, conversational dialogue generation, text-to-speech synthesis, and animated digital avatar — engineered for low-latency interaction with strict quality controls throughout the pipeline.
Deterministic Safety Stack
Multi-layer safety system that operates independently of the underlying language model. Rule-based controls enforce clinical boundaries, validate what the system says about users, and provide a mandatory final checkpoint before any response is delivered. Crisis responses use fixed pre-approved wording only.
Research-Validated Design
Safety controls are validated through formal academic research with trained psychotherapists at Aberystwyth University. Seven phase gates with machine-readable evidence must pass before participant recruitment begins.
Privacy by Design
Three-tier consent model, SQLCipher-encrypted per-participant memory, encrypted research archive stored outside the application, pseudonymous identifiers throughout. Participant data never leaves the secured boundary without explicit consent.
Digital Avatar System
Renderer-neutral avatar bridge with viseme-based lip sync and bounded expression control. Backend selected through measured comparison of UE5 MetaHuman-lite and browser-based TalkingHead — chosen on evidence, not assumption.
Production-Ready Foundation
Built on vLLM for local inference, targeting TensorRT-LLM FP4 runtime for production. Hardware profiles defined from dual RTX 3090 development rigs through to RTX PRO 5000 Blackwell GPUs for dissertation validation, with cloud deployment pathway planned for clinical pilot scale.
From Research to Clinical Deployment
The research instrument runs on local hardware to satisfy ethics requirements and validate the safety stack. The same architecture deploys to AWS for clinical use — providing the GPU compute, scalable storage, and compliance-ready infrastructure needed for healthcare partnerships.
Core Development & Safety Validation
The dialogue engine, multimodal audio pipeline, and deterministic safety stack are built and tested on local GPU hardware. Initial development uses dual RTX 3090 rigs, upgrading to RTX PRO 5000 Blackwell series hardware for the dissertation MVP and full safety validation. Ethics requirements mandate local execution during the research phase — including the Master's dissertation study — so participant data never leaves the secured environment.
Production System & Pilot Readiness
Expanded capabilities: advanced memory, offline analysis tools, read-only researcher dashboards, and full avatar integration. Cloud GPU instances support production benchmarking, TensorRT-LLM conversion, and multi-environment testing ahead of clinical pilot deployment.
Clinical Deployment on AWS
Once safety and ethical gates are cleared, REAA deploys to AWS for clinical pilot use. AWS infrastructure provides the GPU compute for real-time inference, encrypted storage compliant with UK GDPR, and the scalable architecture needed to serve healthcare partners across multiple sites.
Environmental Considerations
Running AI systems has a measurable environmental cost. We take this seriously across the project lifecycle. Development compute is carbon-offset, hardware procurement prioritises energy-efficient configurations, and we actively seek partners and suppliers with verified sustainability commitments.
As the system scales to cloud deployment, we plan to leverage cloud provider sustainability tools — including carbon tracking dashboards, region selection based on renewable energy availability, and right-sized GPU provisioning to avoid idle compute waste. The goal is responsible scale: delivering impact without unnecessary overhead.
Phased Milestones
Phase 1: Core System & Safety Validation
Focus: Complete the dialogue engine, audio pipeline, safety stack, and avatar integration. Validate through academic research with trained psychotherapists.
Gates: Institutional ethics clearance, latency benchmarks, clinician dry-run validation, deterministic safety mapping, seven automated phase gates.
Phase 2: Production Readiness
Following Phase 1 safety validation and ethics clearance
Focus: Advanced memory, offline analysis, researcher dashboards, TensorRT-LLM production runtime, full multimodal capabilities. AWS GPU instances for benchmarking and multi-environment testing.
Quality Benchmarks: Production hardware validation, critic model integration, expanded safety evaluation, pilot readiness assessment.
Phase 3: Clinical Pilot Deployment
Focus: Deploy REAA on AWS infrastructure for clinical pilot use with healthcare partners. Scalable GPU inference, encrypted data storage, multi-site access, and compliance-ready architecture for UK GDPR and NHS data standards.
Quality Benchmarks: All Phase 1 and 2 safety validation complete, secured clinical partnerships, regulatory compliance certification.
Team
Steven Charles
Co-Founder & Lead Developer: System architecture, local inference optimisation, safety pipeline development, full-stack implementation
Alison Mackiewicz
Clinical Co-Founder & Senior Lecturer in Psychology (Aberystwyth University), Registered Psychotherapist: Therapeutic alignment, ethics & compliance coordination, clinician validation
Faisal Rezwan
Technical Co-Founder & Senior Lecturer in Computer Science (Aberystwyth University): System architecture guidance, scalability validation, production deployment strategy