Skip to content

AI Services

DataHub.local uses a hybrid AI strategy: local inference remains available for private or offline workloads, while cloud providers supply higher-capability models for interactive work and automation.

Use case Provider
Interactive chat & coding assistant Claude (Anthropic)
Private/local inference Ollama + Open WebUI
Agents & automated workflows Gemini (free plan) + OpenRouter + Sympozium

Stack Overview

flowchart LR
    subgraph "Inference"
        Claude["Claude\n(Anthropic)"]
        Gemini["Google Gemini\n(free plan)"]
        OpenRouter["OpenRouter\n(multi-model gateway)"]
        Ollama["Ollama\n(local inference)"]
        WebUI["Open WebUI\n(private chat)"]
    end

    subgraph "Automation"
        n8n["n8n\n(AI agents & workflows)"]
        Airflow["Airflow\n(AI pipelines)"]
    end

    User["👤 User"]

    User -->|"chat / coding"| Claude
    User -->|"private chat"| WebUI
    WebUI --> Ollama
    Gemini -->|"OpenAI-compatible API"| n8n
    OpenRouter -->|"OpenAI-compatible API"| n8n
    OpenRouter -->|"OpenAI-compatible API"| Airflow

Services

Claude

llm chat coding-assistant cloud

Primary cloud assistant for interactive use: chat, pair programming, document analysis, and ad-hoc reasoning. Claude is used through Claude.ai and Claude Code (CLI). Models: Claude Sonnet (default), Opus for complex tasks.

Google Gemini

llm agents cloud free-tier

Gemini's free plan handles the bulk of agent workloads in n8n — email summarisation, news digests, document processing, and similar high-volume, low-sensitivity tasks. The OpenAI-compatible endpoint (https://generativelanguage.googleapis.com/v1beta/openai/) means standard n8n AI nodes work without any custom integration. Models used: Gemini 2.0 Flash, Gemini 1.5 Flash.

OpenRouter

llm gateway multi-model cloud

Multi-model gateway used as overflow and fallback when Gemini free-tier limits are hit. The Banana routing tier keeps per-token spend predictable. A single API key, built-in budget caps, and a unified OpenAI-compatible endpoint make it easy to swap between Gemini direct and OpenRouter without changing any workflow logic.

Why OpenRouter alongside direct Gemini?

  • Hard budget caps prevent surprise spend
  • Seamless fallback — same model family, no prompt changes needed
  • Single key covers multiple model families if requirements grow

n8n AI Agents

ai-agents automation llm workflows

n8n hosts all AI agent workflows, combining AI nodes (AI Agent, OpenAI → Gemini/OpenRouter, LangChain, embeddings, vector stores) with real-world integrations. Workflows are stored in datahub-local-workflows/n8n/.

Example workflows:

  • 📧 Email summarisation — classify and summarise incoming emails
  • 📰 News digest — fetch RSS feeds (via CommaFeed), summarise, send daily briefing to Slack
  • 📄 Document processing — extract structured data from PDFs using LLM + regex
  • 🔔 Alert enrichment — take Prometheus alerts, query Loki for logs, generate root cause hypotheses
  • 🏠 Home automation — integrate with Home Assistant events and produce natural-language status reports

Ollama + Open WebUI

Ollama runs on datahublocal-amd-2, the Lenovo Legion GPU worker, and serves qwen3.5:4b for local inference consumed by Sympozium agents. Open WebUI provides a private browser interface. GPU scheduling is provided by the NVIDIA device plugin; workloads must not assume every node has a GPU.

Sympozium

The datahub-local-ai repository deploys Sympozium ensembles into automation. The active ensembles are homelab-ops, homelab-responder, and homelab-reviewer. The repository also contains a Prometheus/Kubernetes MCP server under agents/mcp for bounded fact gathering.

Alvaro Santos Andres