We Built a Weed-ID Assistant Into a Farm Supply Platform
One of our live production platforms is an AI-powered ordering and operations portal for an agricultural supplier. The platform handles product catalog browsing, online ordering, and account management for growers across multiple regions. It's multi-tenant, running on AWS, and used daily.
The feature that changes how growers actually work is the weed-identification assistant.
The Problem It Solves
A grower walks a field and spots something they don't recognize. Or they recognize it and aren't sure whether their current herbicide program covers it. The standard path: take a photo, call the extension office, wait for a callback, cross-reference with the product rep.
That's a slow loop for a problem that has a known answer. Weed identification is pattern recognition against a finite database of species. Herbicide recommendations follow from that identification plus the crop, the application window, and the registered products the supplier carries. The knowledge exists. It just hasn't been at the point of decision.
The assistant closes that gap. The grower describes what they're seeing, or uploads a photo, and gets back a species identification with sourced confirmation plus a recommendation that crosses their herbicide options with what's in their account catalog.
Why It Lives Inside the Platform
We could have built this as a standalone chatbot. We didn't.
The assistant only makes sense inside the account portal because the useful answer isn't just "that's a waterhemp." The useful answer is "that's waterhemp, it's resistant to group 2 and group 9 chemistry, here are the three products in your account that cover it, and the application window for your soybean timing closes in eight days."
That last part: the three products in your account. That specificity requires knowing who the grower is, what they've ordered, and what the supplier actually stocks. A generic chatbot doesn't have any of that. The platform does.
This is the pattern we keep running into. AI features that feel like a bolt-on when described in a meeting become genuinely useful when they pull from the account context surrounding them. The intelligence isn't just the model. It's the model plus the data layer the platform already maintains.
What Sources Actually Means
When we say "sourced recommendations," that's a specific design decision. The assistant doesn't present herbicide guidance as if it's generating it from scratch. It pulls from registered label data and extension publications and surfaces the source so the grower can verify.
That matters in ag because labels are legal documents. A recommendation that can't be traced back to a labeled use is a liability. Building the citation into the output wasn't optional. It was the thing that made the feature something a grower could actually act on rather than a suggestion they'd have to validate before using.
Where This Sits in Production
The weed-ID assistant is one feature in a broader platform that runs continuously for real users. The platform handles ordering workflows, account management, and pricing by region. The assistant gets used when it's needed and ignored when it isn't, same as any other tool.
We've also built out a similar model for Homefront Family Services: a client portal with AI-assisted case management that cuts hours of manual admin per week for their team. Different industry, same pattern: AI embedded in the platform where the work is happening, pulling from the data that surrounds it.
That's what AI in production looks like. Not a demo that impresses in a meeting. A feature a grower uses in a soybean field at 7 a.m. because it's faster than the alternative.