Sohaib A.
Document AI Engineer | OCR, PDF & Drawing Data Extraction with LLMs
I read the documents other tools get quietly wrong. 27,596 scanned annual reports at 27 fields each, every value tied to its source page. Construction plan sets measured to zero delta on all 17 hand-measured rooms. Insurance EOBs cut from about 15 minutes to under a minute per claim. Scanned archives, plan sets, financial filings, EOBs, contracts, Arabic and multilingual documents: the kind generic AI tools read plausibly and get wrong without telling you. Recent results: → Construction SaaS (Australia): reads architectural PDF plan sets and returns rooms, walls and areas. Zero delta on every ground-truth area value, 17 of 17 rooms, 0.96% scale error, about $0.002 a run. → University researcher (US): 27,596 scanned corporate annual reports from 1900 to 1945, 27 fields per document, delivered as validated datasets across five milestones. His review: "the only professional among over twenty contacted who understood historical corporate annual reports." → Healthcare billing (US): EOB extraction across payer formats, from about 15 minutes to under a minute per claim. → Equity research: quarterly and annual filings for 14 listed utilities turned into standardized statements, every figure tied to its source page and checked against the accounting identities before it publishes. → Arabic, Farsi and Russian OCR and translation pipeline, deployed on Google Cloud. How every pipeline works: OCR, an LLM reading pass, then deterministic checks that recompute totals and tie figures across statements, with anything uncertain flagged for a person instead of guessed. You get CSV, Excel or JSON with a source page and a confidence score on every value. How I start: a fixed-price pilot, from $400 for 10 of your real documents, scored against values you already trust. You see the accuracy number before you fund the full build, which is typically $2,500 and up. Send one sample and a sentence on what you need from it, and I'll reply with how I'd approach it and what the pilot would measure. Best fit: teams that have the documents in hand, know what a correct answer looks like, and can sign off on a pilot themselves.