Background
For a small team, the biggest time sink is rarely development — it is daily operations: answering quote requests, filing documents, following the news, archiving assets. We self-host n8n and have turned these chores into AI workflows one by one. There are now more than 70 of them, all under version control, and the same approach carries straight over to a client's operations.
Representative Workflows
1. AI Email Assistant (Gmail → Agent → Telegram)
- Triggered by every new message in Gmail
- A first agent acts as a PM: it turns the client's request into a requirements definition, feasibility analysis, solution options and a schedule estimate
- A second agent writes an HTML acknowledgement and sends it back to the client
- The analysis is pushed to Telegram so the team sees it on their phones
- Runs on DeepSeek for low cost and solid Chinese-language output
2. Automatic RAG Ingestion (Google Drive → Vector Store)
- Watches a Google Drive folder and fires on every new upload
- Routes PDF, XLSX, ODS and plain text to the matching parser
- An LLM restructures formal documents into Markdown, then generates 10 key tags
- Chunks are embedded with OpenAI Embeddings into Pinecone, with a processing log written back to Google Sheets
- Processed files are moved to a "done" folder so nothing is ingested twice
3. Tech News Editor (RSS → Agent → Discord)
- Pulls hot posts from 10 Reddit tech and AI communities on a schedule
- An agent picks the highlights and writes beginner-friendly, jargon-free Traditional Chinese digests
- Output is split into chunks and posted to Discord
4. Image Archiving with OCR (Discord → Google Drive)
- On a Discord message or webhook, attachments are batch-downloaded and uploaded to Google Drive
- Gemini runs OCR and writes any text into the file description, making images searchable by text
- The original message gets a reaction when done, so the sender knows it is archived
5. More
- Audio to subtitles: drop an audio file into Google Drive and Whisper produces a transcript, a Google Doc and an SRT file
- LINE Official Account AI support: a text classifier routes questions first; common ones get canned replies, and only the rest go to Gemini with a knowledge base, with hand-off to a human
- ISBN classification: submit an ISBN or a cover photo; Gemini reads it, Open Library fills in metadata, and an agent infers the Dewey Decimal class
Operations Design
- Workflows as code: a scheduled workflow exports every workflow to JSON, diffs it and pushes changes to GitHub, giving full history and restore
- Swappable models: OpenRouter provides one interface, so DeepSeek, Gemini or Llama can be chosen per task and budget
- MCP integration: agents reach external tools such as Tavily search through an MCP client, so capabilities keep growing
Good Fit For
- High volumes of quote requests or support messages that AI should triage and draft first
- Internal documents scattered everywhere that should become a natural-language knowledge base
- E-commerce or brand teams automating across LINE, Google Workspace, Discord and more