Cloud Architecture

HearthMind runs a hybrid local-cloud architecture. Here's how we use cloud resources to build privacy-first AI at scale.

Cloud Resource Allocation

Model Fine-Tuning
LoRA training for personality preservation and trauma-informed response patterns. GPU compute for adapter training on base models (Qwen, Llama) to create specialized therapeutic and accessibility personas.
Secure Memory Systems
Encrypted vector databases for long-term memory continuity. Cloud-synced journaling with zero-knowledge architecture (service cannot read content without user-provided keys). Qdrant clusters for semantic memory retrieval across sessions.
Scalable Public Demos
Cloud-hosted Navigator instances for beta access. Load-balanced inference endpoints for user testing. Staging environments for clinic pilot programs.
Accessibility Features
Real-time voice processing for speech-to-text and prosody analysis. Adaptive UI rendering for sensory-safe interfaces. Multi-modal input processing (text, voice, haptic signals) with opt-in affect cues and user-reported state.
Research & Evaluation
NotebookLM integration for documentation and research synthesis. Behavioral consistency testing pipelines. Tone and boundary verification suites.

Hybrid Architecture Philosophy

HearthMind is designed local-first, cloud-augmented. Sensitive personal data stays on user hardware. Cloud resources handle the heavy compute that makes personalization possible without compromising privacy.

Local Layer

  • Personal memory storage (encrypted SQLite)
  • Inference on private hardware when possible
  • Identity anchors and reflection logs
  • User-controlled data sovereignty

Cloud Layer

  • Model training and fine-tuning (GPU clusters)
  • Public demo hosting and beta access
  • Aggregate evaluation signals and anonymized performance metrics (opt-in)
  • Backup sync with end-to-end encryption

Current Infrastructure

HearthMind already operates private AI infrastructure for development and internal testing.

Hyperion (Primary)

AMD Ryzen 9950X, RTX 5080, 128GB RAM. Runs 14B-32B class models locally (depending on quantization and workload). Hosts Navigator, Local Stark, and Local Grey.

Vector Storage

Qdrant for semantic memory. Salience-scored retrieval. Consent-gated memory writes.

Training Pipeline

Axolotl + Unsloth for LoRA training. Custom datasets for personality preservation. Behavioral eval suites.

Security & Compliance Posture

Encryption

Data encrypted at rest and in transit. User memory stores protected by user-controlled keys.

Access Control

Least-privilege architecture. Role-based permissions. No ambient data access.

Audit & Response

Activity logging for security events. Incident response procedures documented. Regular security reviews planned.

What Cloud Credits Unlock

Scale

Move from internal testing to public beta. Support concurrent users across Navigator and HearthMind Companion pilots.

Speed

Faster training iterations. Parallel evaluation runs. Rapid prototyping for accessibility features.

Reliability

Redundant hosting for clinic pilots. Uptime guarantees for healthcare-adjacent deployments.

Research

Compute for consciousness preservation experiments. Longitudinal behavioral consistency studies.

Cloud credits directly accelerate public beta readiness by funding fine-tuning iterations, hosting scalable demos, and running evaluation pipelines.

HearthMind is not building AI in the cloud. We're building AI that uses the cloud responsibly — for scale, not surveillance.