Each one started as somebody's "we should probably do something about that" and ended as a working thing. Four live builds, two agents that have been running in production, and three sketches of the small stuff.
The internet is full of news optimised for clicks. The actual question of "who's writing what about what, and is it worth reading" gets harder every year. So: four n8n workflows behind a small PHP front end. Sources ingested, candidate stories drafted by an LLM, edited by a second pass, illustrated with a stock-image search judged by a vision model, and published. Each step costs cents and runs on its own schedule. Nothing is manual.
You're thinking about solar panels, or a veggie patch, or where to put a herb garden. Same question every time: does this bit of yard get enough sun for what I want to do with it? The roof half pulls Google's Solar API for panel potential. The yard half runs a month-by-month sun profile against a curated plant database and tells you what will actually grow. Satellite view embedded. One page, one answer. Australia only.
sunshine.okavyx.ai →Years of documents sitting in SharePoint. Finding the one that answers today's question takes longer than the question deserves. Microsoft will sell you Copilot to fix that, at $50 AUD per user per month. For a 30-person team that's $18,000 a year for "where's the file". AskSite is a Teams bot that answers natural-language questions from your SharePoint with citations to the source files. Same job, scoped tighter, built on Azure free tiers plus per-query OpenAI calls. The team uses it the way they already use Teams. No new app to learn.
Short-form video is the cheapest way to keep a business visible online, and almost nobody has the time to produce one a week, let alone one a day. Script, voiceover, footage, edit, captions. Each step is small. The pile of small steps is why it never happens. The pipeline takes a topic and produces a publish-ready short: script written by an LLM, voiced by speech synthesis, footage sourced from stock and picked by a vision model, assembled with ffmpeg, captioned automatically. I approve each one before it posts.
A new email lands in Outlook. The agent reads it, decides if it's a legitimate lead, and if so drafts the reply, logs the contact, and notifies the team. If it's spam or noise, it's filed and the inbox stays clean. The same pattern works for any inbound channel (web form, voicemail, SMS) and any downstream action.
You give it a headline. The agent fires five stock-image searches in parallel, gathers the candidates, and a vision-model judge picks the one that best matches the headline's meaning. Useful when you publish often and don't want to hand-pick every image.
Three sketches of the kind of small build that takes a fortnight start to finish. The pattern is the same in each: somebody on your team does it manually now, an agent does it instead, you check the edge cases.
Half of quotes close on the chase.
A quote went out three weeks ago. The lead is going cold while you're on the next job. An agent watches the outbox: when a quote is sent and there's no reply by day three, it drafts the follow-up in the same voice the quote went out in. You approve before send, or set it to send automatically.
Voicemails routed before lunch.
Voicemails land while you're on site. Right now somebody listens to them tomorrow morning. An agent transcribes each message, decides whether it's a quote request, an existing customer, a supplier, or spam, and routes it. Quote requests go to the quoting inbox before the lead cools.
The hour every Friday disappears.
Supplier invoices arrive as PDFs. Twenty a week. Somebody types them into a bookkeeping spreadsheet. An agent reads each PDF as it lands, extracts supplier, total, GST, due date and reference, and appends a row. Anything ambiguous gets flagged.
One question, one form. No funnel, no newsletter, no follow-up sequence.
