We read the documents other tools get quietly wrong.
Scanned archives, financial filings and engineering drawings, turned into structured data. Every value is tied to the page it came from and checked before you see it.
What we extracted
Balance sheet, 30 June 1916
| Capital stock | 320,000 |
|---|---|
| Appropriated surplus | 7,729 |
| Profit and loss surplus | 171,945 |
| Printed total | 499,674 |
Lines add up to the printed total, which matches total assets.
- Company
- Pittsburgh, Cincinnati, Chicago & St. Louis Railway Co.
- Incorporated
- 1 October 1890
- Miles operated
- 1,488.98
Moody's Manual of Investments, Steam Railroads, 1917, p. 599. Public domain scan, University of Toronto, via the Internet Archive.
Selected work
Measured on client documents, against values checked by hand.
| Documents | What we delivered | Measured result |
|---|---|---|
| Historical annual reportsUS university research team | 27,596 scanned corporate reports from 1900 to 1945, turned into a 26-field research dataset: financial lines, auditors, officers and directors. | Delivered in five milestones. The client's spot checks against the scans found almost no errors. |
| Architectural plan setsConstruction estimating software, Australia | A plan reader inside a live product: rooms, walls and areas, drawn back onto the plan for the builder to confirm, with a feedback loop that learns from corrections. | 17 of 17 hand-measured rooms matched. 0.96% scale error. |
| Insurance statementsUS healthcare billing | Explanation-of-benefits extraction across payer layouts, with a plain-English action for the biller on each claim. | About 15 minutes per claim, down to under one. |
| Stock exchange filingsFinancial services firm, Philippines | Standardised statements for 14 listed utilities. Statements located by content, every figure linked to its source page. | Accounting identities checked on every filing. The same input gives byte-identical output. |
| Engineering drawingsPower plant archive, 42,000 drawings | A reading pipeline for title blocks, tables and identifiers across scanned and vector drawings. | Title blocks read 96% right on a hand-checked set. |
“I communicated with over twenty professionals before Sohaib and Sohaib was the first person that understood historic corporate annual reports.”
How a project runs
You see the accuracy before you pay for the build. Research and library teams get the sample free; companies start with a paid pilot.
Sample
You send a few real pages and say which values you need. We return them extracted, each value tied to its page.
Measure
We score the sample against values you already trust and show you field-level accuracy, including what we flagged instead of guessing.
Build
We run the full set. The model reads; our code checks totals, cross-references and identities. Anything that fails goes to a person, never into your data unmarked.
Research
Read, then verify: financial data from 1878–1917 railroad manuals
An open benchmark and dataset from five public-domain volumes of Poor's and Moody's manuals: officers, directors, income accounts, balance sheets and funded debt. We publish a hand-keyed gold set, the full pipeline, and the share of errors it catches before they reach your data.
In progress. Code, data and paper due December 2026. Ask to be notified.
About
Olympus Labs is an engineering company registered in Pakistan, founded by Sohaib Ali. It is small on purpose: the person who scopes your project is the person who builds it.
Write to hello@olympuslabs.ink with one sample page and a sentence on what you need from it. You'll get a reply with how we'd approach it and what a pilot would measure.