Axiara

Costing and quotation bring together several kinds of information: an official price baseline, experience in historical documents, and changing market prices. Treating them as interchangeable can make a tidy answer difficult to trust.

I want Axiara to help with the work around a quote — gathering information, finding a relevant reference, filling a batch, and checking anomalies — while keeping the authority of each source visible.

The core decision is to separate the official baseline, learned historical information, and market information into three data layers. Historical documents and crawled prices can inform the work, but do not become official prices just because an agent found them.

The official main layer is maintained manually. The learned and market layers have different write boundaries. I see this as part of the product model: deciding who may change a price is more important than making every step automatic.

Archive, query, batch quote, and review are capability and permission boundaries. They are not a mandatory sequence of screens, and a task does not have to pass through all four.

• Archive: maintain versioned official prices and organize historical or market references under their respective permissions.
• Query: look up a single item’s official cost, market range, and process notes.
• Batch quote: work through Excel or BOM items with templates and constraints.
• Review: compare a quote against the baseline and market references, and surface anomalies.

FastAPI exposes the service, and LangGraph provides agent orchestration. The repository uses SQLite alongside file-first CSV and YAML data. The point of the separation is to make the source and permitted use of information explicit, rather than leaving that distinction only in a prompt.

A useful example is a market observation that differs from the official baseline. It may be relevant to a review, but it should not silently rewrite the baseline. Gathering evidence and authorizing a change are separate actions.

The service, separated data layers, and business modes give Axiara a starting structure. The next questions belong to concrete tasks: looking up one item, preparing a batch, and explaining why a quoted value deserves review.

The direction I want to keep testing is whether someone can trace a result back to its inputs and understand which decision still belongs to them. An agent earns a place in the workflow by making that clearer.