Case study Three clients · Documents into working tools

They already had
the information.
They just could not use it.

A coatings manufacturer with hundreds of technical data sheets in Word. A supplier with 696 products and no usable photography. A family with a box of scanned policy documents. Same shape of problem three times: the answers exist, in a format nobody can search, sort or check.

Morsun Coating Systems Acuserve
Plus one private family office
The technical data sheet catalogue
Data sheets catalogued
205
Products imaged
696
205
technical sheets, searchable
696
products through the review tool
617
verified by a human, one by one
12
correction rounds on the catalogue
Manan Mehta with the AI Crew team in the Mumbai office
From our office
Real people. Real office.

This is the crew in Mumbai that builds what you see on this site. No stock team, no AI-generated faces.

The AI Crew team at work
Who you work with
Manan Mehta and the crew
arrow_forward
01 · The pattern

A document is a record. A tool is an answer.

Almost every business we meet is sitting on the information it needs. It is in a folder of Word files, a shoebox of scans, a spreadsheet somebody maintains by hand. The information is not missing. It is just in a shape that can only be read one page at a time, by a person who already knows where to look.

folder_open
Findable only by the one who filed it

Which makes the whole archive a dependency on one person's memory.

compare_arrows
Impossible to compare

Two products, two documents, and no way to put the specifications side by side.

fact_check
Never checked as a whole

Errors live for years, because nobody has ever seen all of it at once.

Why this is an AI job now

Because the hard part was never the database. It was reading a few hundred inconsistent documents accurately enough to trust the result. That used to be a data-entry project with a budget nobody would approve. Now it is a week, and the real work moves to verification.

02 · Three of them

Word files, product photos, and a box of scans.

The TDS catalogue
Coatings manufacturer
205 data sheets, one catalogue

Every technical data sheet, parsed out of their own documents into a searchable catalogue with a clean PDF per product. A customer finds a product code in seconds instead of emailing to ask.

The image review tool
Facilities supplier
696 products, reviewed not guessed

Catalogue imagery at the scale of a whole product master, with a review tool so a human approves or flags each one. Verified, to review, flagged and finished, as running counts.

inventory_2
A private family office
Scanned policies, one live register

A folder of scanned policy documents with no text layer, turned into a passcode-protected register of what is held, what it pays and when each one renews. Client anonymous by request.

account_treeHow the three tools work

Documents in. Something you can search, sort and check out.

Reading the documents takes an afternoon. Being able to stand behind every line is the project.

Comes in
description205 Word data sheetsA coatings manufacturer
inventory_2696 products, no photosA facilities supplier
document_scannerA box of scanned policiesA family office
account_tree

Extract, then verify

Reading is the easy half. Checking is the job.

  • check_circleDocument parsing and OCR on scans
  • check_circleA review tool with four states
  • check_circleCorrection rounds, logged and closed
Goes out
searchA searchable catalogueA clean PDF per product
photo_library617 verified imagesEach approved by a person
lockA live policy registerRenewals, behind a passcode
Client
folder_open01Hand over the archive

Word files, a product master, or a folder of scans with no text layer.

AI
document_scanner02Read every document

Parsed and read into structured fields. This part takes an afternoon.

Human
fact_check03Verify each one

Four states and a count. 617 verified means a person looked at 617.

Client
rate_review04Correction rounds

Their words win and their order wins. Every round logged.

Build
picture_as_pdf05Publish the tool

A catalogue with a clean PDF each, or a passcode-protected register.

Client
search06The archive answers

A product code found in seconds instead of an email to ask.

Built withdescriptionDocument parsingdocument_scannerOCR on scanspicture_as_pdfGenerated PDFsfact_checkHuman review toolinghistoryCorrection round logs
03 · The part everyone underestimates

Extraction is the easy half. Verification is the project.

Reading the documents takes an afternoon. Being able to stand behind every line of the result takes twelve rounds with the person who wrote them. That is not a failure of the tool, it is what the job actually is.

Verify against what is live, never the local copy

Twice on the catalogue, the build on the machine was ahead of or behind production. Checking the wrong one wastes a whole round.

Hard rule
The client's words win

A product code that looks like a typo is sometimes the real code. Their proper nouns are correct until they tell us otherwise.

Judgement
Match their order, not a better one

She checks the sequence of rows against her own document. Helpfully re-sorting a table is a correction round we caused.

Faithful
A human approves each one

On the imagery job that meant a review tool with four states and a count against each. 617 verified means a person looked at 617 of them.

Counted
Print the discrepancy, do not fix it silently

One data sheet's own quantities table contradicts its own heading. We published it verbatim and raised it, because it is their document to correct.

Honest
Case studies · built by the crew

More case studies. All live.

All case studies
04 · What changed

The archive started answering questions.

Before
  • closeSpecifications emailed on request, one at a time
  • closeHundreds of products with no usable photography
  • closeScanned documents nobody could search
  • closeErrors invisible because nothing was ever seen whole
After
  • check205 sheets, searchable, with a clean PDF each
  • check696 products imaged, 617 human-verified
  • checkA live register with renewal dates, behind a passcode
  • checkTwelve rounds of corrections, closed and logged
Document parsingOCR on scansGenerated PDFsHuman review toolingCorrection round logs