Software for Models and Agents
The model worked in a notebook. Why does it change when the API reorders fields, another request joins the batch, or a job reads an old artifact? Build the software boundaries around those failures, then use data structures to keep retrieved evidence and workflow dependencies intact.
Keeping prediction code dependable
The same values can reach a model in a different order. A procedure that rescales readings can also reuse statistics from the wrong training run. We’ll write procedures that make those assumptions explicit.
- 01Ready to exploreThe Right Numbers in the Wrong Columns
An API changes its field order. Keep the trained model attached to what each feature means.
- 02Ready to exploreA Fitted Object Should Remember Its Training
Separate fitting from prediction so a new request cannot quietly redefine the model’s reference.
Structures for agent workloads
A retrieved passage needs to stay attached to the score that selected it. A training job needs its input job to finish first. Those problems give us reasons to use different data structures.
- 03Ready to exploreKeep the Useful Candidates Without Losing Their IDs
Use indices and a bounded heap to select retrieval candidates while preserving ties and identities.
- 04Ready to exploreWhich Jobs Can Start Before the Model Is Ready?
Represent a training workflow as dependencies, detect cycles, and separate readiness from completion.