The Document Extraction Pipeline
Turns one kind of document your business receives, such as supplier invoices or delivery dockets, into records in your database or files. A person checks anything the AI flags, and it runs on your own server and AI account.
Someone on your team opens each supplier invoice, finds the supplier, the number, the date and the total, and types them into another system. The typing adds up to hours each week, and a slip in one total can sit unnoticed until the books don't balance.
We build a pipeline that collects those documents from an email folder, a folder on your server, an upload page or your own system, and has AI read up to 15 details plus the line items from each one. Every value goes through checks you choose, such as totals that add up or an invoice number you haven't seen before. Your staff review anything that fails a check or that the AI marked as unsure, side by side with the document, before it reaches your database, your files, a webhook or your Admin Panel.
Before handover we run it on 10 to 20 of your own documents and show you how often it matched your answers, and which mistakes the checks caught. It runs on your server and your own OpenAI, Anthropic, Gemini or Mistral account, so you pay the provider yourself and we never see your key.
What's included
One kind of document, such as supplier invoices or delivery dockets, from any number of senders, with up to 15 details and an optional table of line items read from each, set out in a written plan you approve before we build
Documents arrive the way they do today: an email folder, a folder on your server, an upload page or a webhook from your own system. Repeats are skipped, and files it cannot read are set aside without an AI call
Up to eight checks you choose, such as totals that add up, a valid ABN or an invoice number not seen before, so a misread value is stopped before it reaches your records
A review screen where your staff see the document beside its details, correct anything flagged and approve it, with every correction logged. Start by checking everything, and switch to flagged-only once the results earn your trust
Approved records land in new tables in your database, JSON and CSV files on your server, a webhook, or an Admin Panel we built for you
A measured run on 10 to 20 of your own documents with a report on how often it matched your answers, one tuning pass, a second report, then one round of changes
Runs on your Docker host or Linux server and your own AI account, with a monthly spending cap, alerts when something fails, the source code, and a plain-English handover that records where your documents go
Before you order
Have these ready so we can start the moment you order. The delivery clock only runs while we have them.
A Docker host behind an HTTPS reverse proxy, or a Linux server that already serves PHP 8.2 or later over HTTPS and has poppler-utils installed, that we can reach over SSH (you add the deploy key we send)
The kind of document, the details you want from it, and three to five examples to show us on the planning call
Access to where documents arrive (an email folder that accepts a password or app password, or a folder on the server) and to where records should go (database details, a folder or a webhook address)
An account with billing turned on at OpenAI, Anthropic, Google Gemini or Mistral (your Administrator adds the key in the pipeline; we never ask for it)
One person to approve the plan and later supply 10 to 20 test documents with the correct answers, plus up to two Administrators
How it works
Book it — pay the fixed price and tell us the target environment. No scoping call first.
We deliver — an experienced specialist runs our standard workflow for this exact task. You can watch progress and message them in the portal.
Hand-off — you get the deliverables, the documentation, a rollback snapshot, and a walkthrough if you want one.
Who does it
A salaried Support Resort specialist with more than ten years in the field, working from a runbook we've been refining since 2003 — supervised by senior management, never a marketplace contractor. You'll see who is assigned and can message them about this task directly in your portal.
FAQ
Ask about this task
Our assistant knows the scope document and runbook behind this service. Ask what's covered, what isn't, or whether it's the right task for your situation.
Answers come from our published scope documents. Prefer a person? Switch to the "Email us" tab — a senior manager replies personally, usually the same weekday.
Related services
The MCP or API Integration
One of your systems, or your AI assistants, connected to a platform you already pay for, such as Xero, HubSpot, Shopify, Stripe or Microsoft 365, so nobody types the same record twice.
The Admin Panel
A proper back office for your staff, built on the data you already have: up to ten sections they can search and edit, with a record of who changed what.
The AI Portal
Your best people's AI prompts, turned into simple forms the whole team can use, on your own AI accounts and under a monthly spending cap for each person.
The AI Developer Day
One full day of a senior AI developer working your list: LLM features shipped, prompts and retrieval measured against real cases instead of guessed at, and token bills brought back down.