AI document validation, built for production

Turn a pile of documents into a decision you can stand behind.

Interpret reads an entire dossier — PDF, Word, Excel, CSV — checks it against your own rules, flags what's missing or inconsistent, and produces a traceable report a person reviews and signs off. Fast like AI. Defensible like a human decision.

interpret.app/reports/1042
Validation report
Mortgage application — Alex Whitfield
In review
Policy: Loan application review Version: v4 Documents: 5 Date: May 28, 2026
Needs review 5 pass · 1 missing · 1 n/a 7 checks across 5 documents
Government-issued photo ID present Pass 100%

Driver's licence on file for Alex Whitfield.

Photo ID is valid (not expired) Pass 99%

Expiry date Apr 12, 2029 — well in the future.

Name consistent across documents across set Pass 98%

"Alex Whitfield" matches on the ID, the application form and the bank statement.

Proof of income attached Pass 100%
Spouse co-signature present across set Needs review 96%

The applicant selected "Married" on the application, but the spouse's signature on page 3 is missing — additional information required before approval.

Bank statement is recent Pass
AI: Needs review Reviewer: Pass “Statement is dated within the last 90 days — confirmed acceptable.” R. Mehta · May 28, 2026, 2:44 PM
Marketing consent form N/A — not required
Whole-dossier
many documents judged together, with per-document traceability.
Cross-document
catch inconsistencies no single-file check can find.
Human-in-the-loop
review, override with a reason, finalize. The AI never has the last word.
Traceable
every check, its evidence and who decided — exported to PDF/JSON/CSV.

Built for teams where a document error is a real risk

The real problem

A demo impresses in minutes.
Production breaks in the details.

The hard part was never reading one clean document. It's judging a messy dossier of many — and being able to explain the verdict afterwards.

Real dossiers are messy

Multiple formats and versions, missing pieces, documents that contradict each other, endless exceptions. Clean demo files hide all of it.

Rules live in people's heads

The criteria that decide admissibility are scattered across emails, habits and one expert who's on vacation — so they're applied inconsistently.

You have to justify the call

A black box that says "rejected" isn't enough. You need who checked what, against which rule, with what evidence — and to defend it later.

How it works

From a drop zone to a signed-off report

One simple flow. Upload a dossier, and Interpret takes it from raw documents to a decision you can export and defend.

Upload

Drop the whole dossier — many files, any format.

Read any format

PDFs & images rendered; Word/Excel/CSV read as text.

Apply your policy

Your rules, in plain language — required, optional, scoped.

Cross-check

Consistency across documents, approved-list lookups, anomalies.

Human review

Annotate, override with a reason, finalize.

Report & export

A clean report, exported to PDF, JSON or CSV.

Embed anywhere

Run this pipeline wherever your work already happens

Interpret is API-first. Trigger a validation from your website, your app, a chat bot, an email inbox or your document store — and get back a structured, traceable verdict you can act on. No rip-and-replace; it slots into what you already run.

Documents arrive from…

🌐
Your website & web app
Drop-in upload or a few lines of API.
✉️
Email inbox
Give a policy its own address — built in.
🗂️
SharePoint / shared drive
Watch a folder, pull the dossier.
Interpret pipeline

Validate · Review · Trace

POST /api/validate
{ "policy": "loan-application",
  "documents": [ application.pdf, id.jpg, … ] }
200 OK
{ "overall": "needs_review",
  "report_url": "…/reports/1042",
  "results": [ 7 findings ] }

Results flow out to…

🔌
REST API & webhooks
Verdict + report JSON, pushed on completion.
💬
Slack / Teams bot
Post the result to the channel that owns it.
🧩
CRM / ERP / GED
Write the status back to the system of record.

It can live as a hosted app, a panel inside your product, a bot, or a headless service behind your own UI — the validation engine is the same everywhere.

See it in action

A guided tour of the product

Real screens from Interpret. It plays itself — or tap a tab to jump.

interpret.app/validate
Checks
🛡Loan application review
6 rules · identity, income & signatures
KYC onboarding
5 rules
Insurance claim intake
4 rules
⬆ Drop documents to validate
PDF, Word, Excel, CSV or images · Loan application review · 6 rules
Application — Alex Whitfield5 docsneeds review
Application — Jordan Romero4 docscompleted
Application — Sara Patel5 docsrunning…
Validation report
Application — Jordan Romero
Policy: Loan application review v4Documents: 4
Complete5 pass · 1 n/a
Government-issued photo ID presentPass100%
Proof of income attachedPass100%
All required signatures presentacross setPass
Marketing consent formN/A
Proof of income
Draft

The applicant attached a signed employer offer letter instead of a pay stub, so the AI flagged it for a human to confirm.

AI: Needs reviewReviewer: Pass “Signed offer letter stating salary is acceptable proof of income for this product.”
VERDICT Pass · AI Fail Needs review ✓ Finalize report
Loan application reviewv4
01Name consistent across documents Across set Required
02Accepted ID type Each doc Required
The value must bein Accepted ID types ▾ Match: Smart
03Spouse co-signature when married Across set Required
LibraryDocument typesReference lists
Accepted ID document types4 items
Passport
Driver's licence
National ID card
Residence permit

One value per line — paste a column straight from a spreadsheet. A rule can require a document value to be in or not in the list.

Everything in the box

Built for the messy, high-stakes middle

📑

One case, many documents

Drop a whole dossier and validate it as a single case. Every finding can point to the exact document it came from.

🗂

Reads any format

PDFs and images are rendered page-by-page; Word, Excel and CSV are read as text. Mix them freely in one dossier.

⚖️

Policies & rules, versioned

Write rules in plain language with accept/fail criteria. Every change is snapshotted — view history and restore any version.

Across the set

Rules that reason across the whole dossier

Choose each rule's reach: check each document on its own, or across the entire set — "the applicant's name must match on every form," "if form A says Yes, form B must be attached." Catch the inconsistencies no single-file tool can.

  • Per-document, any-document, or set-wide reach
  • Names which documents it compared, in evidence
  • Conditional logic between forms
Applicant name consistentacross setFail

The application form shows "Daniel R. Cohen"; the bank statement shows "Daniel Cohan" — the names don't match.

application.pdf bank_statement.pdf
Accepted ID typereference listFail

The ID provided is a "Foreign voter card", which is not on the approved list "Accepted ID document types".

Reference data

Check values against approved lists

Eligible models, approved vendors, valid codes. Keep them as editable lists in the Library — paste a column from a spreadsheet — and point a rule at one: the value must be in or not in it, exact or smart match. No code.

  • Spreadsheet-style editing, CSV paste
  • Exact (auditable) or smart (tolerant) matching
  • Edit the list → the verdict updates next run
Human-in-the-loop

A person reviews, and signs off

The report opens as a draft. A reviewer adds notes, overrides a verdict with a required reason, and finalizes. The AI's original answer is never erased — the report shows "AI said X → you set Y, because…" with a timestamp. Reopen to amend; it's all logged.

  • Per-finding notes & overrides, attributed
  • Overall verdict recomputes from your edits
  • Finalize / reopen with full audit history
Passport valid
AI: FailReviewer: Pass “Renewal receipt attached; valid through 2029.” R. Mehta · May 28, 2026
📄

Reports built to be read

Verdict first, problems surfaced first, evidence on drill-down. A finished document, not a debug dump — on screen and on paper.

⬇️

Export anywhere

One-click PDF that mirrors the screen exactly, plus structured JSON/CSV to feed the systems you already run.

✉️

Email intake

Give a policy its own inbox; a dossier emailed in is validated automatically, and the reply carries the result.

0
document formats read
0%
of decisions traceable
documents per dossier
0
export formats (PDF/JSON/CSV)

Use cases

Wherever a document decision must hold up

Interpret is generic — the rules and lists you configure make it yours. Explore three ready-made setups, or browse the dossiers teams run through it:

Receive · validate · prioritize

Application completeness

Detect missing pieces and signatures, validate completeness, decide Complete / Incomplete / Needs more info.

Programs · subsidies

Grant & subsidy applications

Check eligibility against criteria and approved lists, with a reviewer who signs off each file.

Compare · flag · justify

Financial document validation

Reconcile amounts, dates and statuses across forms, statements and invoices; flag the inconsistencies.

Control · trace · audit

Regulatory & compliance

Verify the expected documents are present, coherent and conformant — with an auditable trail.

Read · structure · check

Contract & legal review

Surface clauses, dates, parties and signatures across long documents and check them against your rules.

Onboard · verify

KYC / identity onboarding

Validate ID documents and proof-of-address as a set, with consistency checks across them.

What Interpret is — and isn't

Serious tools say what they're not

Not just OCR

It doesn't only read text — it ties information to your rules, to its source, and to a verdict.

Not a chatbot

A structured process with a checklist verdict — not a conversation you have to re-interpret.

Not a black box

Every result is explained, evidenced, correctable and exportable. You can defend it.

Not a replacement for experts

It clears the repetitive checks so your people spend their judgment where it matters.

FAQ

Questions, answered

What kinds of documents can it read?

PDFs and images (rendered page-by-page) plus Word, Excel and CSV (read as text). You can mix every format inside one dossier and validate them together.

Do I need to write code to make rules?

No. Rules are written in plain language with accept/fail criteria, and configured with simple toggles — required/optional, each-doc/any-doc/across-set reach, and an optional approved-list check.

Can a human override the AI?

Yes. A reviewer can change any verdict with a required reason and finalize the report. The AI's original answer is preserved and shown next to the human decision, with a timestamp and full audit history.

Is it auditable?

Every check, its evidence, the documents it touched, and every human note/override/finalize are recorded — and exportable to PDF, JSON or CSV. Policies are versioned, so you always know which rules produced a given verdict.

Which AI models does it use?

Interpret calls vision/LLM models through OpenRouter, so the model is configurable. Your rules and reference lists steer it; the human review keeps the final say with a person.

See Interpret on your own dossier

Bring one real process — the documents, the rules, the exceptions. We'll set it up and show you a traceable verdict end to end.