Architecture

The Three-Repo Trident

KynticAI is not a monolith. It is three purpose-built repositories that work together: an open-source core, a SaaS control plane, and an enterprise fortress.

Scout

Open-Source Core

The public data-plane engine. Domain entities, selector execution, context snapshots, GraphQL/REST API, React admin console. MIT-licenced.

universalcontextlayer

Cloud

SaaS Control Plane

Account management, billing (Stripe), licensing (RSA-signed), data-plane registration, downloads, support cases, and the mini CRM. Never touches raw customer data.

universalcontextlayer-cloud

Fortress

Enterprise On-Prem

Private connectors (SQL Server, PostgreSQL, REST/CRM, email), governance, compliance exports, identity (OIDC/SCIM/SAML), credential vaults, Helm charts, air-gapped deployment.

universalcontextlayer-enterprise

Data Flow: Source to Consumer

Data never moves. KynticAI reads metadata where it lives and builds semantic context in place.

1

Sources

SQL databases, CRM, ERP, flat files, APIs — data stays where it lives

2

DataSource Registration

Connectors register metadata without moving raw data

3

Selector Execution

5 mapping kinds: DirectField, StringToEnum, Threshold, WeightedScoring, Formula

4

Context Facts

Semantic units with confidence scores, provenance chains, and freshness

5

Context Snapshots

Point-in-time assemblies for any entity, consumed by AI agents

6

Consumers

AI models, dashboards, REST/GraphQL APIs, webhooks, SDK clients

Two Enterprise Sub-Tiers

Without-AI Tier

Semantic context layer only. Connects to existing databases, builds context facts with provenance, serves via API. No AI model required — pure data infrastructure.

  • Connector marketplace
  • Selector engine
  • Context snapshots
  • GraphQL + REST API
  • Provenance audit trail
  • Role-based access

With-AI Tier

Everything in Without-AI, plus embedded AI model support. Llama, Mistral, or customer’s own model runs on-prem. The context layer feeds the model; the model enriches the context.

  • Embedded Llama/Mistral
  • On-prem inference
  • Discovery Agent
  • Self-improving flywheel
  • Synthetic industry agents
  • AI-safe packaging

Rust Engine Modules

The performance-critical path is written in Rust for memory safety, zero-cost abstractions, and sub-millisecond latency.

ucl-ingest

High-throughput metadata ingestion from connectors

ucl-selector

Selector execution engine with 5 mapping kinds

ucl-context

Context fact assembly with confidence scoring

ucl-provenance

Full audit trail from fact back to source record

ucl-synthesis

Markdown dossier generation with token budgeting

ucl-gateway

API gateway with rate limiting and auth

Architecture Intention

MetricLegacyKynticAI
Selector executionBatch reports and manual ETLLow-latency selector path
Context snapshot assemblyStatic dashboard exportOn-demand governed context
Integration modelCentralised copy-first pipelineRead-only customer data-plane access
Operational footprintLarge platform dependencyScoped services per deployment
Change controlSlow project release cyclePilot-first validation path
Deployment postureGeneric cloud dependencyCustomer-owned data plane

See What Connects to KynticAI