Before · Re-typed by hand
The data existed. It was locked inside PDFs from several different vendors.
Every site had a report from an external vendor, and each vendor wrote its report differently: different column names, different field ordering, different ways of labelling the same item. Before anyone could act on a report, a team member had to find the right entry, read the row, and verify the status against the source PDF.
One person kept a parallel spreadsheet current by re-typing the reports, page by page. Nobody trusted it enough to act on it without re-reading the source PDF.
Off-the-shelf tools lost the nesting between header rows and detail rows, or the rows that crossed page breaks. None could tell which vendor format a given PDF used. No accuracy or time figure for the manual spreadsheet is on the record.
Build · First cut in 8 weeks
Two passes, then a person checks it beside the source.
The system reads each PDF twice: once to pull out every table cell with its layout, then one row at a time, with a cloud language model extracting the regulated fields the team needs.
Each row is checked against the team’s existing vocabulary, fixed when the obvious things are wrong, and recovered when the parser missed a section. The reviewer interface shows the value on the left and the highlighted region of the source page on the right.
The pipeline runs on AWS Bedrock in Melbourne, so the source PDFs and every extracted value stay inside Australia, and the model provider contract excludes the team’s data from any future model training.
First production cut within eight weeks. Refined sprint by sprint since.
Read the layout. Every table cell comes out with its place on the page, including cells that span rows or columns.
Every table cell comes out with its place on the page, including cells that span rows or columns.
A cloud language model extracts the fields the team needs from each row on its own.
Each value is checked against the team’s vocabulary; obvious errors are fixed and missed sections recovered.
The value on the left, the highlighted region of the source page on the right.
Parsing Table cells read with their layout, spanning cells includedModel A cloud language model on AWS Bedrock in Melbourne; a local model as backupRecords Stored with their source link, protected against two people overwriting a rowExport In the format the team’s existing system expectsProvenance Every value carries its source page and highlighted region
In production · Tracked per job
90% accuracy and $1.65 a document, tracked per job.
Accuracy is tracked per job, so the team sees the score for the document in front of them, not just the average.
Outcome figures and architectural choices are real and verified against the production rollout. Client identity and specific product names have been generalised so the pattern reads across verticals.
$1.65 · 31 min per document
production cost and wall-clock per long-form document, typically 150 to 250 pages
One long document, 150 to 250 pages
Production cost$1.65
Wall-clock31 min
- Checked by
- Verified against the production rollout.
No earlier figure on the record, and the team’s cost and turnaround targets are not stated.
90%
extraction accuracy across the reference document set the team validated against
- Window
- Tracked per job
- Checked by
- Verified against the production rollout.
No earlier figure on the record.
Adoption · Week two
The source link was the adoption result, not the accuracy.
Reading a whole table in one go reads beautifully on the happy path, and fails silently the moment the source has a merged cell, a page break, or a row the vendor flagged differently than expected. Reading one row at a time is more work. Every row either comes back or it does not, and the system can tell you which.
The accuracy score was the engineering result. The source link was the adoption result: the team will not act on data they cannot verify. We would build the source link first next time.
Every value
links back to its source PDF page with a highlighted region
How the system behaves, not a measured outcome.
“The first week, two people kept opening the source PDFs anyway, just to check. By the end of the second week nobody was opening them. That was the day we knew it was working.”
Have a stack of reports nobody fully trusts?
A 45-minute audit. We look at the documents your team relies on, where the data needs to end up, and what the actual bottleneck is between the two. You get a written follow-up within 48 hours whether or not we are the right fit.
