EPOCH.
An AI nervous system.
Seven architectural layers modeled on biological autonomy — afferent pathways for sensing, central nervous system for processing, efferent pathways for action, proprioception for feedback, autonomic governance for safety. 132 MCP tools across 22 production servers. Local-first by default, constitutional by design.
Seven layers, modeled on a nervous system.
Biological nervous systems are distributed, autonomous, and accountable to a body. EPOCH applies the same shape to AI infrastructure. Each layer is a real, replaceable component in the live system — not a diagram.
L1 · Afferent Pathways
Sensory inputs. Voice (Whisper STT), vision (reCamera + Hailo face detection), IoT sensors (Ecowitt weather, Home Assistant), location (Traccar), calendar, network telemetry, music state.
L2 · Central Nervous System
Processing centers. LangGraph multi-agent orchestration, three-tier LLM routing, confidence-aware autonomy, deterministic reasoning (temperature 0, seeded).
L3 · Memory Systems
Three tiers: Redis (hot cache), Memori PostgreSQL (working memory), OpenSearch (vector / semantic). Gibson Memory is the canonical truth source.
L4 · Nervous System Wiring
Communication pathways. MCP HTTP Gateway v6 with 132 whitelisted tools across 22 Pattern A production servers. Redis event bus. Event-driven inter-service coordination.
L5 · Efferent Pathways
Action systems. TTS (F5-TTS, Kokoro), MCP tool calls, SMS/phone via Asterisk + Telnyx, physical actuators on Reachy, Home Assistant scenes, music systems.
L6 · Proprioception
Feedback loops. Outcome tracking per action. LLM-based scoring of severity and confidence. Telemetry taxonomy (accept / intercept / dispatch / rejected).
L7 · Autonomic Nervous System
Governance & safety. Constitutional principle enforcement. Schema, semantic, and safety validation layers. Idempotent replay and step persistence. Professional-grade output as a guarantee.
Fourteen constitutional principles.
The Constitution is the contract every part of EPOCH agrees to. Seven core tenets define what the platform is. Seven guiding tenets define how it operates.
Core tenets
- 1. Deterministic Intelligence
- "AI systems must produce consistent, reproducible outcomes." Temperature 0, seeded, decision trees not randomness.
- 2. Validated Output Integrity
- "Every AI output must be validated against desired outcomes before action." Schema, semantic, and safety layers gate every action.
- 3. AI-First Nervous System
- "Infrastructure designed around AI capabilities from inception, not retrofitted." Semantic over keyword. Embedding routers over regex.
- 4. Intelligent LLM Routing
- "Right model for right task — optimize for outcome, not convenience." FAST for latency, SMART for reasoning, DEEP for long context.
- 5. Human-Indistinguishable Synthesis
- "TTS must be natural, emotional, and contextually appropriate." MOS > 4.0/5.0, blind-test indistinguishable, ≥7 emotions, <500 ms first token.
- 6. Confidence-Aware Autonomy
- "AI must calibrate autonomy based on confidence — high confidence enables action, low confidence requests human judgment."
- 7. Professional-Grade Output
- "Every artifact EPOCH produces must meet a standard worthy of the platform — mediocrity is a bug."
Guiding tenets
- 8. Privacy-First, Local-First
- Zero cloud dependencies for core functionality. On-premise inference for all sensitive data.
- 9. Graceful Degradation
- Multi-tier fallbacks (GB10 → xAI → cloud-emergency). Offline-capable. Fail-safe defaults.
- 10. Measurable Intelligence
- Benchmarked performance. A/B testing. User-feedback loops. Explainable decisions.
- 11. Unified Memory & Context
- Gibson Memory as canonical truth source. Persistent context across sessions and modalities.
- 12. Developer Experience Excellence
- Simple APIs. Observability (Prometheus, structured logs). Rapid iteration. Comprehensive docs.
- 13. Atomic & Idempotent Components
- "Every component should be replaceable, removable, or duplicable without breaking its neighbors."
- 14. Durable Orchestration
- "Multi-step workflows must survive crashes, restarts, and partial failures without requiring full replay."
Four-tier intelligence.
One model can't do everything well. EPOCH routes each request to the right tier — fast for reflexes, smart for reasoning, deep for long context, edge for fully air-gapped. 256 GB pooled VRAM across three NVIDIA accelerators in the home server, plus a Hailo-8 NPU for the standalone product.
FAST · ~172 tok/s
Qwen 3.5 35B-A3B (MoE) on RTX 5090, 32 GB. Wake word, simple factuals, lights, music play. Latency-critical.
SMART · ~195 tok/s
gpt-oss 120B on RTX Pro 6000, 96 GB. Tool orchestration, advisories, reports. 4.75× faster than GB10 — the workhorse.
DEEP · ~41 tok/s
gpt-oss 120B on DGX Spark GB10, 128 GB unified Blackwell. Long-context, transcripts, document deep-dives.
EDGE · 40 TOPS
Pi 5 + Hailo-8 NPU. The path that powers Aria Standalone — air-gapped capable, voice and vision never leave the unit.
EPOCH is running, not pitched.
MCP HTTP Gateway v6 is deployed. 132 tools operational across 22 Pattern A production servers. Aria runs daily on Reachy, AudioSocket telephony, and a Pi5 standalone unit. The Constitution is enforced by code, not by intent.