Denys Shkola
EmailCV (PDF)

Project 11AI engineering

n8n agents on a Pi

Six self-hosted n8n workflows on a Raspberry Pi 5 that use only local models (Ollama with qwen3:8b), Qdrant memory and a private SearXNG: a market-news digest, a daily AI-engineering lesson in Notion, a retrieval chat agent and a weekly recap.

Code on GitHub
Status
Public repo · MIT
Role
Solo: workflows, infrastructure, tests and operations
Period
Aug 2026
Stack
n8n · Ollama (qwen3:8b, nomic-embed-text) · Qdrant · SearXNG · Docker Compose · Python tests

The problem

I wanted useful daily agents that run on a small home server with local models only: no cloud LLM keys, no data leaving the machine, and output that can be trusted without reading every run.

What I built

What I built
Workflow Trigger Output
Market News Digest 07:00 and 18:00 e-mail with cited bullets and a watchlist
AI Engineering Learning Feed 05:30 up to 3 concept pages and a daily briefing in Notion
Ask the Library n8n chat answers from my concept library first, the web only when it has no answer
AI Weekly Recap Sunday 18:00 the week’s concepts by category, with no model calls

Two support workflows send failure alerts and serve the Notion backlog to the chat agent.

How it works

  • Grounded selection. The scout may pick only numbered candidates, never URLs. A parser validates the picks against the fetched items and recovers complete arrays from truncated JSON. Tests run the exact Code-node JavaScript.
  • Prompt budget. On the Pi 5, qwen3:8b ingests about 7.7 tok/s and generates about 2.0 tok/s. A prompt-budget failure, where 55 candidates made a 3,641-token prompt and 320 output tokens truncated the JSON, was fixed with a 28-candidate cap and compact, source-indexed JSON.
  • Semantic dedupe. Qdrant similarity scores were measured (identical 1.0, reworded 0.816, unrelated 0.454) to set the threshold, scoped per workflow and to a 7-day window.
  • Operations. Images pinned by tag and digest, a nightly systemd backup with 30-day retention, no-write validation clones and an import tool that wires credentials and cross-workflow references.

Results

count
6n8n workflows: 2 production pipelines plus 4 support workflows
operations
≈ 2.0 tok/sqwen3:8b generation on the Pi 5 (about 7.7 tok/s ingest)
operations
3 min 43 sone full on-device validation run (Aug 2026, a single development run)

A full validation run took 3 min 43 s on the device in August 2026, a single development run. Live-data testing shaped the design: the model now only adds to the lexicon’s list of stories, so a small model can never remove a genuine one.