Layer 1
Public Edge
Cloudflare Pages (SSG build público) + Cloudflare Tunnel + Access / Zero Trust. Separación público/privado con build filter y _redirects edge.
Cloudflare Pages Cloudflare Tunnel Cloudflare Access Zero Trust Traefik Proxy
AI-LAB BLUEPRINT
Arquitectura del Ecosistema
Plataforma de ingeniería con runtime evidence-bound, gobernanza enterprise, knowledge base RAG, marketplace digital twin y observabilidad integral. Arquitectura en capas desde el edge público hasta el backend de inferencia.
Capas de la plataforma
↓
Layer 2
Documentation & Portal
Astro con Starlight, Tailwind, Mermaid y React. Dos builds: público (ai-lab.labrazahome.com) y privado (blog-ai-lab.labrazahome.com).
Astro Starlight Tailwind CSS Mermaid React/Recharts
↓
Layer 3
AI Runtime Gateway
Entrypoint único de chat. Clasificación de ruta, aplicación de perfil cognitivo, inyección de memoria, SLO enforcement y adaptive concurrency.
OpenAI Gateway (Python) Route Classifier Profile Manager Memory Injector SLO Manager
↓
Layer 4
Inference Backend
LM Studio con 3 modelos activos: llama-3.1-8b (operacional), qwen2.5-coder-14b (coding/report), nomic-embed-text-v1.5 (embedding). GPU RX9070.
LM Studio llama.cpp RX9070 GPU OpenAI-compatible API
↓
Layer 5
Runtime APIs
Router API (status, perfiles, replay), Live API (estado, embeddings), Hermes Status Endpoint (:8095), precision endpoints.
Router API Live API Hermes Status Runtime Precision
↓
Layer 6
Hermes Enterprise Core
Capa de governance declarativo: SOUL, Capability Registry, Operator Registry, Hook Registry, MCP Registry, Dynamic Governance. 185 tests.
SOUL Capability Registry Operator Registry Hook Registry MCP Registry Dynamic Governance
↓
Layer 7
AnythingLLM Knowledge Base
RAG enterprise con 1304 vectores en 7 workspaces. Embedder multilingual e5-small. Chunk size 800, overlap 100. Validación 100%.
AnythingLLM Qdrant e5-small Embedder Workspace Manager
↓
Layer 8
Observability Stack
Prometheus (source of truth), Grafana (15 dashboards), Loki (logs). 100+ métricas ailab_*, 19 alertas activas, SLO enforcement.
Prometheus Grafana Loki node_exporter cadvisor
↓
Layer 9
GitNexus Code Intelligence
Cognición estructural del codebase. 20K+ símbolos, 32K+ relaciones, 300 execution flows. Impact analysis, context, rename, detect_changes.
GitNexus Index Knowledge Graph Impact Analysis Process Flows
Modelos activos
Operacional
llama-3.1-8b-instruct
Minimal, greetings, observe, light prompts
Coding + Report
qwen2.5-coder-14b-instruct
Coding, report, architecture, reasoning, creative
Embedding
nomic-embed-text-v1.5
Semantic recall, vectores Qdrant
Checkpoints clave
CP-HERMES-ENTERPRISE-CORE-01
6 componentes enterprise, 185 tests
CP-ANYTHINGLLM-ENTERPRISE-04
Knowledge Base, 1304 vectores
CP-36B-RUNTIME-PRECISION-MODE
Runtime precision mode
CP-SLO-ENFORCEMENT-01
SLO enforcement framework