In 2017, “chatbot” did not mean a large language model. It meant rules you wrote by hand.

I was an agriculture undergraduate who had started teaching myself to code, and the gap I most wanted to close was the distance between a farmer looking at a sick plant and the agricultural knowledge that could name what was wrong. The knowledge existed — much of it sat in Taiwan’s open government datasets — but it was not in a form anyone could just ask.

So my team and I built one of the first Mandarin-language chatbots for plant disease diagnosis. A farmer described the symptoms in plain language; a rule-based natural-language pipeline mapped that description onto the government’s agricultural disease data and returned a likely cause. No neural network, no pretrained model — just careful, unglamorous work connecting how people talk about a sick plant to how the data records disease.

We entered it in the Open Data Innovative Application Contest, run by Taiwan’s Ministry of Economic Affairs and Council of Agriculture, pitched it to a panel of venture investors, and came away with NTD$410,000 in prize funding.

It reads like a period piece now — this was years before conversational AI became ordinary. But it is also where a pattern I still follow began: take a real problem someone actually has, reach for open data and whatever computational tools fit, and build the smallest thing that helps. The tools have changed completely since. The instinct has not.