Nexi is an adaptive AI with a decision pipeline, not a chat endpoint.
Most AI systems accept input and generate output. Nexi runs every request through a 15-step pipeline — from intent classification through policy alignment, scoring, and execution dispatch. Each step is observable, auditable, and subject to governance.
XNCH's first system. Deployed on a dedicated inference node with GPU-accelerated inference, a YAML-defined character, and continuous learning.
Decision Pipeline
Every session passes through 12 sequential stages before an action is dispatched. Two additional stages handle outcome processing and learning.
Input Validation
Request is parsed into a typed SessionContext — actor identity, raw input, priority, idempotency key, system state version.
Intent Interpretation
Two-stage classifier: rule-based pre-filter (regex patterns, confidence 1.0) with LLM fallback. Detects injection attacks. Flags ambiguous inputs for clarification.
Context Manifest Loading
XNCH memory is queried for relevant episodes, learned patterns, and scoped policy references matching the intent class and entity type.
Weight Config Retrieval
Active decision weights for the intent class are loaded from XNCH governance. Each weight (policy, outcome, risk, context fit) tunes how options are scored.
Option Generation
LiteLLM proxy generates N candidate action options via structured JSON output. Falls back through vLLM, llama.cpp, then rule-based templates if inference is unavailable.
Policy Alignment Filter
All options are sent to XNCH for parallel dry-run policy evaluation. BLOCKed options are dropped. MODIFY verdicts replace the action spec.
Scoring
Each surviving option receives a weighted composite score across four dimensions: policy alignment, expected outcome (from pattern history), risk (reversibility, sensitivity, side effects), and context fit.
Outcome Simulation
The top-2 riskiest options receive forward state projection. A risk penalty is applied if pattern history indicates violation risk.
Selection
The highest-scoring option is selected. Confidence is measured as the spread between best and second-best. Low confidence triggers escalation.
Plan Compilation
The selected action spec is validated and compiled into a DAG node. Currently single-node; designed for multi-step plan expansion.
Authoritative Verdict
XNCH re-evaluates the selected action and issues a signed JWT execution token. BLOCK verdicts are written to the immutable decision ledger.
Execution Dispatch
The signed payload — execution token, action spec, trace context — is dispatched to the execution runner.
Character System
Nexi is not a generic language model wrapper. It has a defined persona, loaded from a YAML character file at runtime and compiled into every system prompt.
Persona Definition
Non-Negotiables
Hard Rules
Disagreement Protocol
Proactivity Engine
Nexi does not wait for input. A background engine monitors system health and behavior patterns, surfacing observations without being asked.
Stale Patterns
Failure rate > 60% in any pattern cluster
Surfaces underperforming patterns for review
Consolidation Stale
> 25 hours since last learning loop run
Alerts that the learning pipeline may need attention
Inference Down
vLLM health check fails
Triggers fallback chain notification
Learning Loop Silence
Pattern extractor silent > 7 hours
Raises an observation for investigation
Maximum 5 proactive messages per day. Never during deep work. Never nags the same thing twice.
Model Architecture
Nexi routes every request through a configurable inference chain with automatic fallback. No single point of failure for intelligence generation.
Development Status
Nexi is in active development. Core behavioral frameworks, the decision pipeline, and character system are operational. Public access is not yet available.
We are working with a small group of partners to refine the system before broader release.