A mortgage file can contain several versions of the same fact. An application states one income figure, supporting documents imply another, and a later attachment changes the picture again. Finding a number is one task. Establishing which number a reviewer should use, and why, is a different one.
Parsewise builds document intelligence for that second problem. It brings information from a collection of files into a shared record, with source references and a review process for people who need to validate the result. Its public YC profile describes multi-document processing that can identify relationships and contradictions across files. [1]
A document problem shaped by operational experience
Founded in 2024, Parsewise joined YC’s Spring 2025 batch. CEO Maximilian Hofer brings a machine-learning research and consulting background; CTO Greg Csegzi previously worked at Palantir. Their product is aimed at work in which documents are inputs to an operational decision, rather than the end of the workflow. [1]
That distinction matters for financial and insurance teams. A fluent answer can still leave the reviewer with the original job of checking every attachment. A useful system needs to make the coverage of the file and the evidence for each conclusion easier to inspect.
One record across the document collection
The Parsewise Data Engine describes building a persistent, structured representation of a task and its documents. Its stated approach is to process every page, connect related facts and preserve page-level sources, rather than rely only on a small set of retrieved passages. The company also describes using reviews and corrections to improve subsequent processing. These are product-design claims, not a guarantee that every conclusion is correct. [2]
Its published claims example shows why coverage matters: a change in a case may be expressed differently in an expert report, an adjuster’s note and an email. Looking only for one phrase can miss the connection. Parsewise illustrates how the engine brings that evidence together; the example should be understood as a product scenario, not a reported customer result. [2]
Mortgage review: reconcile the file before the next handoff
A Google Cloud customer story describes Parsewise’s use in mortgage-document review. It reports reducing a review that could take up to two hours to about 15 minutes. The story emphasizes resolving conflicting information and pointing reviewers back to the original scanned evidence. That is a provider-published account of the company’s deployment experience, rather than a universal processing-time promise. [3]
The practical workflow is to assemble the file, reconcile the relevant values and let a person inspect exceptions before the result moves onward. For a lending team, the useful question is whether an analyst can understand a mismatch without reopening and searching every document. The lending decision, the institution’s policies and the responsibility for resolving uncertain evidence remain with its staff.
The same customer story introduces Navi as a conversational way for business users to work with the document information. Its role is to connect a question to evidence already organized by the engine. That can make a follow-up investigation more accessible to the reviewer who owns the file. [3]
What the public OfficeQA result establishes
Csegzi also pointed us to Parsewise’s OfficeQA write-up. The company tested questions over a large collection of US Treasury bulletins, a setting that stresses cross-document reasoning and numerical evidence. Its post reports 78 correct answers out of 133 against the original answer key, or 58.65%. [4]
The post separately reports 69.92% after proposed revisions to 15 reference answers. That second figure uses the company’s revised answers and should not be treated as the same result against the original key. The distinction is useful to preserve when interpreting the headline. The write-up makes its reasoning available so readers can examine the disputed cases. [4]
For an operations leader, the relevant connection is the review process itself: can someone trace an answer through the documents, challenge an assumption and determine whether the source or the generated answer needs correction? A benchmark can provide evidence about a specific test. It does not settle the accuracy of a different organization’s files or approval rules.
Where Parsewise fits on SOTA2
| Product & task | Position | Comparison scope |
|---|---|---|
| Parsewise PlatformExtract Document DataFinancial Services & Insurance | #9 | 39 companies · 52 candidates 10 ranked products |
| Parsewise PlatformExtract Document DataHealthcare & Life Sciences | #3 | 16 companies · 17 candidates 10 ranked products |
| Parsewise PlatformMortgage Loan File ReviewFinancial Services & Insurance | #3 | 10 companies · 12 candidates 10 ranked products |
| Parsewise PlatformVerify DocumentsFinancial Services & Insurance | #7 | 23 companies · 28 candidates 10 ranked products |
| Parsewise PlatformVerify DocumentsHealthcare & Life Sciences | #6 | 9 companies · 10 candidates 7 ranked products |
What we would look for in a pilot
We would start with a completed set of mortgage or insurance files whose exceptions have already been reviewed. Include conflicting values, revised documents and evidence that is difficult to read. Compare both the resolved fields and the source references with the reviewers’ existing conclusions. Record missed evidence separately from incorrect interpretation.
The operational measure is the time and effort needed to produce a reviewable file. Track how many exceptions still require a full manual search, whether corrections carry through to related values, and whether the next reviewer can understand the record without repeating the work.
The SOTA2 view
Parsewise is relevant to teams whose document work depends on completeness and human validation. Its appeal is the connection between a shared record, cross-document reasoning and inspectable evidence. Mortgage-file preparation is one concrete place to test that connection, while the final decision remains with the institution.