The N is not a letter — it is an infinity, folded.
The cyan triangle is the fold — the point where infinite becomes operable.
AI-Inference First. The Neural Spine wires the LLM to your precision instruments, workflows, documents, and personas on a single nerve. Every feature answers one question: in the next iteration, how many human decision points does this remove?
The architecture is the product. Chat Engine routes tools. Workflow Engine executes DAGs. Event Bus chains them. 5 circuit breakers hold the line when any subsystem stalls.
- Chat Engine
- Workflow DAG
- Event Bus
- BullMQ + Redis
- LibSQL · per-user
- MCP custom servers
Three layers, one nerve.
A user types in chat or a sensor pushes a frame. The spine routes, executes, and audits. Memory persists what survives the gate. That is the whole product.
Where your team types or your sensor pushes. Same chat for sourcing audits and field disease scans. Same API for any precision instrument.
Where decisions get made. Workflow Engine runs your DAG, Event Bus chains workflows, five circuit breakers keep one stalled subsystem from wedging the rest.
Three-gate Gatekeeper validates before anything persists. Below threshold goes to human review. Per-user namespace, AES-256-GCM at rest. Your data never crosses tenants.
No subsystem is unbreakable.
Every external dependency lives inside a breaker. Open → fast-fail. Recovered → half-open probe. A subsystem without a breaker doesn't earn its way to production.
LLM
Provider failover. Multi-pod safe. Backs off + recovers automatically.
MCP
Per-server isolation. One bad MCP server cannot starve the others.
Event Bus
Wildcard subscriber storm protection. Slow listener never wedges the hot path.
Audit
Persist queue + guarded writes. Causal chain never lost on DB stall.
Quota
Per-workspace caps. One workspace cannot exhaust shared LLM budget.
What gets remembered, you approve.
Memory persistence is the slowest possible drift. We refuse to write anything until three gates pass. Below threshold goes to Gatekeeper queue for human approval. Memory you can trust.
Extraction gate
Is there an extractable claim? Filters small-talk, noise, transient state.
Classification gate
Preference / Fact / Episode / Relationship — refuses if type is unclear.
Confidence gate
Score against existing memory + conflict detection. Below threshold → human review.
Persisted
All three gates passed. Memory enters long-term store, scoped by workspace + user. Conflict detection alerts on contradictions.
Pick the model that fits your audit posture.
A curated default ships out of the box. Swap in any OpenAI-compatible endpoint at configuration time — your model, your endpoint, your audit log. Provider neutrality is the rule; the supply-chain detail lives on the Security page.
OpenAI Built-in
Anthropic Built-in
Claude Built-in
Groq Built-in
Google Gemini Built-in
OpenRouter Built-in
Ollama Built-in · self-hostedCustom provider surfaces in the UI as Custom. No origin is celebrated or suppressed at the surface.
Every LLM call carries a provider tag into the audit log. Customers can verify which endpoint actually answered.
Inference stays inside the deployment. No third-party hop, no data egress, no vendor lock-in.
Architect the spine. Personas come for free.
A platform is what it can refuse to break. We architect for the bad day, not the demo day.