Services / AI Document Analysis
AI that reads your entire document archive, and shows its work.
Some questions can only be answered by reading everything: every contract, every claim file, every scanned letter in the box nobody has opened since 2009. JTPCK builds AI document analysis systems that read collections of any size, find the passages that matter, and link each result to the exact page it came from.
Proven at scale
We built a pipeline that searched 4.35 million handwritten pages from the Dutch East India Company archive, plus centuries of newspapers, in a single overnight indexing run on one GPU. It screened tens of thousands of passages for a few dollars and surfaced findings missing from the specialist catalogues, including an uncatalogued 1812 meteorite fall and volcanic eruptions absent from the Smithsonian's global list.
Read the case study for the architecture, the costs, and the results.
Problems this solves
- Finding every document that mentions a person, product, clause, incident, or pattern, not just the ones that use the right keyword
- Searching scanned PDFs, faxes, and old records where text recognition is full of errors
- Screening thousands of files to decide which ones a professional actually needs to read
- Pulling dates, names, amounts, and parties out of unstructured text into a spreadsheet or database
- Answering "is it in there anywhere?" with enough confidence to act on a no
How we build it
Search by meaning. We index your documents with embeddings so the system finds passages that describe what you are looking for, even when they use different words, abbreviations, or misspellings.
Screen cheaply, read carefully. A small, fast model makes the first narrow yes-or-no call on every candidate. Only what passes goes to a more capable model for close reading and extraction. This model cascade is what makes reading everything affordable.
Test against known answers. Before we trust the system, we make it find documents you already know are there. If it can find those, its empty results mean something.
Cite the source. Every result links to the original page, so your team verifies findings instead of taking a model's word for them. A person stays in charge of the conclusions.
Where this works well
- Litigation, discovery, and case file review for law firms
- Due diligence on contracts, leases, and vendor agreements
- Audits and record reviews for accounting firms and finance teams
- Insurance claim files, compliance reviews, and regulatory responses
- Research literature reviews and medical records summaries
- Corporate, municipal, nonprofit, and historical archives
Document analysis or RAG?
A RAG AI system answers questions: someone asks, and it retrieves the few most relevant passages to answer from. Document analysis is exhaustive: it sweeps the whole collection to find, classify, or extract every instance of something. Many projects need both, and we build them on the same index.
For confidential collections, the indexing and screening can run on hardware in your own office, so documents never leave the building.
Common questions
How many documents can AI analyze?
There is no practical ceiling for most businesses. Our archive project indexed 5.7 million passages from 4.35 million pages overnight on a single GPU. The real constraints are document quality and how precisely the question can be defined, and we work through both during scoping.
Can AI read scanned documents and handwriting?
Yes, once they have been through text recognition. The resulting text is often messy, which is why we search by meaning instead of exact keywords and check results against the original page images.
How do we know the AI did not miss anything?
We test the system against documents you already know are in the collection before running the real search. If it reliably finds those, you can trust its results, including when it reports that something is not there.
How much does AI document analysis cost?
It depends on the size of the collection and the questions you need answered. Because a small screening model does most of the reading, the computing cost is usually modest. Every build is quoted as a fixed price before work starts.
Can our documents stay private?
Yes. We can run indexing and screening on hardware you control, and use business-grade model accounts with no training on your data for the steps that need a larger model.
Get started
Have more documents than people to read them?
Tell us what you need to find and where it lives. We will tell you whether AI can find it, and what it will take.
Talk to us about document analysis