The problem
A die-casting operation consumes materials continuously and documents that consumption on paper. Orders and receipts pile up as scans and PDFs, get filed, and are functionally invisible from that point on — the information exists, but answering "what did we spend on this material last quarter, and what will we spend next quarter?" means someone opening documents one at a time.
The ERP system knows what was ordered in the abstract. It does not know what the documents say unless a person types it in, which is slow, error-prone, and the first thing to fall behind when the floor is busy.
Forecasting is the visible deliverable, but it is downstream of a data problem. You cannot project spend on a material you only have a filing cabinet's worth of evidence about.
Document ingestion
The first half of the work is getting documents into a form a system can reason about. AWS OCR handles extraction — material orders and receipts are analysed automatically, pulling out line items, quantities, prices and dates rather than treating the page as a flat block of text.
Managed OCR was the right call over a self-hosted model here. The documents are commercial forms with recurring but supplier-specific layouts; the value is in reliable structured extraction at low operational overhead, not in owning a model. The engineering that matters sits around it — normalising output across suppliers whose invoices agree on nothing, matching extracted line items against known materials, and flagging what needs a human rather than silently guessing.
Forecasting
With orders and receipts structured and accumulating, resource and financial projection becomes tractable. The tools cover both sides of the same question:
- Resource forecasting — consumption patterns per material, so procurement can anticipate demand instead of reacting to a shortage.
- Financial forecasting — projected spend built from actual historical document data rather than from budget assumptions, which means the forecast and the ledger are describing the same reality.
Transparency
The deliverable people actually use is the overview. Extracted and projected data is surfaced through visualisations that make spend and consumption legible at a glance — the point being that a manager should be able to see where resources are going without first knowing which question to ask.
That framing drove the design. A forecast that only a finance specialist can interpret changes nothing about how decisions get made; the same numbers presented as a clear trend get looked at, argued with, and acted on.
Why it was worth doing
The satisfying part is how ordinary the inputs are. No novel model, no exotic infrastructure — just a document pile that everyone had accepted as unreadable, turned into a live view of the business by taking automated extraction seriously and then presenting the result plainly.
Specifics are limited — this is commercial work for a client's internal systems.