FIELD NOTES / 001 / THE BEGINNING

A new quest for machine learning.

Useful machine learning starts long before a model makes a prediction. That’s the part of the adventure we want to help with.

A business question sounds simple: what will our customers order next month? Answering it means understanding accounts, joining orders to products, setting a time boundary, choosing a baseline, and checking whether a forecast is actually useful. There are a lot of small decisions between a question and a dependable answer.

GraphForge is an early research and development project exploring a way to connect those decisions: verifiable AI agents for machine learning pipelines over relational business data.

Give the quest a clear objective.

A task should say what success means before a model gets involved. Which entities matter? What information was available at prediction time? What will we measure, and what budget do we have? Keeping those choices explicit gives every later step something concrete to work toward.

Bring the right tools.

We’re exploring an architecture in which agents coordinate specialized tools, while code processes the underlying records. An agent can reason over computed findings and references to artifacts. It doesn’t need to carry an entire dataset through a conversation.

That leaves room for many approaches: an existing query, a seasonal baseline, a tree model, or a suitable foundation model. The best option should earn its place through validation on the task at hand.

Keep a useful save file.

A pipeline should leave a trail: its code, input references, assumptions, checks, and results. When the business changes, that trail helps distinguish a modeling problem from a broken join or a changed definition. And when something works, it can become reusable execution rather than another prompt to start over.

If this sounds like your kind of problem, stick around. We’ll use this space to share the ideas and experiments behind GraphForge.