Background

Over the past two months I’ve followed an online course: Solve It. More than a product, this is a considered approach and method for working with AI, coined dialogue engineering.

I cannot share the course material, at least and perhaps not yet. But I can convey that one of the lesson was showing how to use the rather complicated and extensive Cloudflare python API to perform cloud engineering, something that was attempted without success some time ago “pre-AI”. The key aspect of the exercise was to create the right context for the AI to competently propose code and steps, where the AI alone would have spouted things of dubious quality.

Closer to home

One line of work I have is short-term streamflow forecasting. For a few years we’ve had python bindings for swift2 openly published, so recently trained AIs (which happens to be Claude Sonnet 3.5) may well have seen this in training.

So, can it competently propose to help, giving a bit of context (not reproduced here), with a basic modelling setup?

Thank you for providing the documentation link. I can now see this is CSIRO's SWIFT framework for hydrological modeling. Looking at the documentation, I can help you create a code snippet for a GR4J model.

Would you like me to start with a basic setup that:

    - Creates a GR4J model instance
    - Sets up a single subcatchment
    - Uses sample climate data

Let's start with importing the necessary modules and creating a GR4J model instance. Based on the documentation, here's the first step:

from swift.swift import Swift
from swift.swift_types import *

# Initialize SWIFT
swift = Swift()

# Create a GR4J model instance
gr4j = swift.create_model('gr4j')

Yeah nah, definitely an hallucination.

Planning a line of work

End goal

Can we enable a hydrologist, trained but not previously exposed to the swift2 package, to set up a correct, executable semi-distributed hydrologic model?
How far can we push the definition of complicated calibration workflows with minimal coding by the hydrologist?

Steps

I should eat my own dog food and use upcoming modelling needs on this from to test whether this is of use to me as well. After all, I do forget stuff and am adverse to copy/paste and not a particularly fast typer.

Once I have a context I find satisfying, identify a new user not familiar (or not overly familiar) with the python API, and set successive tasks. It is a dual research question, as one user’s approach to using an AI buddy may well vary from another. This is testing both humans and a contextualised AI, really.