Skip to main content
EigenPal is organized around a loop. You build something that automates a task, prove it is correct, run it, watch it on real inputs, and feed what you learn back in. Each primitive in the platform exists to serve one stage of that loop. The example below extracts a bank statement, but the same loop fits any workflow.

​
Build

Compose a workflow, a directed graph of typed steps that parse, extract, transform, and act on a document. For open-ended tasks that do not fit a fixed graph, train a Git-backed agent instead. Either way, the definition is a file you can read and version, separate from the engine that runs it. 
eigenpal init workflow extract-statement
eigenpal workflow push

​
Evaluate

Attach a dataset of example documents with their expected output, and evaluators that define what correct means for this document type. Run an experiment across the whole dataset and tune the workflow until it clears a pass threshold. Quality stops being an opinion and becomes a number. 
eigenpal workflow evaluators push <workflow-id>
eigenpal workflow experiment run <workflow-id> --wait

​
Deploy

Run the workflow from the dashboard, the CLI, or the API. The engine is stateful and queue-based, so a request returns a run id and the work completes reliably even under load. Every run is stored with its inputs, outputs, model versions, and scores. 
eigenpal run workflows.extract-statement -F document=@statement.pdf

​
Observe

Because every run is kept, you can inspect exactly what happened: the per-step trace, the output, and where it came from. This is where you catch the misses that no test anticipated. 
eigenpal runs trace <run-id>

​
Improve

When a real document exposes a wrong answer, promote that run into the evaluation dataset with the corrected output. Re-run the experiment, and the fix is now permanent: the same mistake cannot reappear without a test catching it. Future versions are validated against every issue you have ever seen. This is the closed loop, production issues become test cases. It is the whole point, and it is walked through end to end in Eval-first development.