Document AIIn production since June 2026

Documents in.
Decisions out.

Phone photos from your customers, scans, PDFs, emails, handwriting. A BrilliantFlux system checks the shot on the spot, reads the fields you need, tests them against your rules and files the result into your CRM, ERP or DMS. People only see the cases that need a person.

10document types in one production deployment
99.8%page classification accuracy, own CPU model
~13 msper page on CPU, no GPU required
RU · EN · DEplus 20+ languages on request
status: reading…
InputIMG_4471.heic · passport, data page
Extractedconfidence
Checked & routedyour rules
Processed in 0.9 s · runs on your servers or EU tenant · 0 bytes to external APIsClick any field to see where it came from · passport is a fictional specimen
Ask the document:

What teams automate with it.

The samples above are examples. The same pipeline handles whatever arrives from your customers' phones, your inbox, scanner or shared drive.

IN

Collecting documents from customers

The case that runs in production today. You send a link, the customer photographs each required document on the phone. Every shot is checked before it is accepted, fields are read and confirmed by the customer, and your staff see a queue of complete cases instead of an inbox of attachments. Missing or rejected pages go back to the customer with one click, and the loop closes without email ping-pong. Insurance claims and dependent eligibility, KYC and onboarding, visa and civil-status paperwork, tenant screening, loan applications.

in phone camera, PDF → out case file, CRM, staff queue in production · 10 document types
What the customer sees
  1. Link, language, consent
  2. List of required documents, computed from a few questions
  3. Camera, instant "too dark / cut off / retake"
  4. Extracted fields to confirm
  5. Submit, then a message if something needs a retake
AP

Incoming invoices

Supplier, totals, VAT, IBAN, PO matching, approval limits. Duplicates and changed bank details flagged before payment.

in PDF, scan, email → out ERP, accounting
KYC

Identity documents

Passports, IDs, residence permits, proof of address. MRZ reading with checksum validation, expiry checks, consistency across the documents of one case.

in phone photo, scan → out CRM, case file RU · EN · DE in production
§

Contracts & clauses

Parties, term, renewal, fees, liability and non-standard clauses. Reminders before auto-renewal. Search across the whole archive.

in DOCX, scan → out DMS, calendar, legal queue
✎

Forms & handwriting

Intake forms, questionnaires, delivery notes, inspection sheets. Handwritten fields with confidence scores, low ones go to a person.

in paper, photo → out database, HIS, ticket
▦

Tables & reports

Bank statements, price lists, lab results, multi-page tables that OCR breaks. Table structure recognition is our published specialty.

in PDF, scan → out spreadsheet, database CVPR 2026 paper
✉

Inbox triage

Classify incoming email and attachments, extract what matters, route to the right person or system, answer the routine ones.

in mailbox → out CRM, helpdesk, ERP

Your documents are not on this list? They rarely are. Describe the flow on a call and we tell you which parts automate and which do not.

Book a call

How it works.

Four stages, each one auditable. Every field knows which pixels it came from.

01 · Capture

Get a clean page

Guided photo capture, scan or PDF intake. Quality checks catch blur, poor light and cut-off pages before they become bad data.

02 · Understand

Classify and read

Page type, layout, tables, stamps, signatures, handwriting. Own models trained on your document set, run on CPU or your GPU.

type: invoicelayout: 2-coltable: 3×7 ✓
03 · Extract & check

Fields with confidence

Key-value extraction, table cells, cross-document consistency. Your business rules run on every result. Low confidence goes to a human.

auto ✓
04 · Act

Write where you work

Results land in your CRM, ERP, DMS or spreadsheet with an audit trail. Corrections your team makes become training signal.

JSONERPCRMDMS

What is checked before a page is accepted.

Each check runs on every photo in the production platform. The customer sees the verdict in a second and retakes on the spot. Nothing here is a promise, it is what runs today.

▣
Document found in the frameOwn detector locates the page. No document, no upload: "place it on a plain, contrasting background".YOLO11 · ONNX · CPU
☼
BrightnessToo dark or washed out is rejected before OCR ever runs on it.all document groups
◈
SharpnessMotion blur and out-of-focus shots, the most common reason a page cannot be read.12 of 13 groups
⊞
Margins & cut-off edgesThe whole page must be inside the frame. A clipped MRZ line is a useless MRZ line.where the layout requires it
Aa
Text readabilityEnough edge detail in the text region for reliable recognition.text-heavy groups
px
Minimum resolution & duplicatesThe detected page must be large enough in pixels. The same photo uploaded twice is dropped by content hash.all groups

You decide where the human sits.

Every extracted field carries a confidence score. You set the threshold per field, per document type. Above it, straight through. Below it, a person confirms in a two-click review screen. Drag the threshold to see the trade-off on a typical month of invoices.

  • 1
    No silent errors. A wrong IBAN never reaches payment unseen; it is either confident or reviewed.
  • 2
    Reviews shrink over time. Each confirmation trains the model on your documents. Month three needs fewer people than month one.
  • 3
    Audit trail by default. Who confirmed what, when, from which pixels. Useful for your auditor and your EU AI Act file.
confidence threshold2 400 invoices / month
63%straight through
37%human review
44 hreview time / month

Illustrative curve. The real one depends on your documents.

In production, not in a slide deck.

One platform used every day since June 2026. Client named on the call, not on the website.

In production · legal-services company, DE/DK

Document collection, verification and extraction platform

The client's customers photograph passports, IDs, certificates and forms on their phones in three languages. The system computes which documents each case needs, guides the photo, checks quality, reads the document, asks the customer to confirm, and hands a clean structured case to staff. Staff accept or reject page by page; rejected pages go back to the customer and the case returns when fixed. Built end to end by BrilliantFlux: models, backend, web app, CI/CD, GDPR-aware consent and retention.

10document types
100+client cases processed
99.8%page classification, top-1
0.995mAP50, page detector
YOLO11 detectorsONNX on CPUMRZ + key-value extractionFastAPIReactRU · EN · DE24 e2e test chains
1
Guided capturephone camera or PDF, live quality feedback
2
Auto-classificationmulti-page PDFs split and typed, ~13 ms per page
3
Extraction + user confirmationthe person who knows the answer checks it once
4
Staff review queueaccept or reject per page, request changes in one click
5
Structured case outinto the client's workflow, fully traceable

What it does not do.

Every first call asks this. A short honest list makes the rest of the page believable.

It does not replace the reviewer.

The system prepares, checks and flags. A person makes the accept or reject decision on anything below your confidence threshold. That is a feature, and your auditor agrees.

It does not guess on unreadable pages.

A blurred or cut-off photo is sent back for a retake at the moment it is taken. No "best effort" fields end up in your database.

It does not take payments or sign contracts.

Payment and e-signature stay in the tools you already have. We write the status into them, we do not replace them.

It does not chat about your archive.

Questions in free text over thousands of documents are a different system with a local LLM. That is the Private AI direction, and the two combine well.

Want the numbers on your own documents instead of ours? Bring three real ones to the call - accuracy is measured on your documents in the pilot, before we commit to a figure.

Book a call

PIX-TAB · CVPR 2026

Research

Pixel-precise table structure recognition with speculative decoding, co-authored by the founder. This is the table-reading know-how behind the "Tables & reports" use case.

Read the paper →

ICDAR 2025 · 1st place

Competition

First place in the 2025 international document-analysis competition, second in 2024 (multi-font OCR). The benchmarks the field measures itself by.

Need it fully private?

Prototype

The same pipeline runs inside your network with a local LLM for questions over the archive. See the Private AI direction.

Private AI →

Asked on every first call

Our documents are messy: stamps, handwriting, three languages. Does it still work?+

That is the normal case, not the edge case. The production deployment above runs on phone photos of IDs and certificates in Russian, English and German. Accuracy is measured on your documents before we commit to numbers.

How is this different from off-the-shelf OCR or a cloud document API?+

Generic OCR gives you text. We give you the fields you need, checked against your rules, in your system, with a confidence score and a review screen for the rest. And it can run on your own hardware, which cloud APIs cannot.

Do we need a GPU?+

Not for extraction. The production models run on CPU at about 13 ms per page. A GPU is only needed if you add a local LLM for free-text questions over the archive.

How long until the first documents flow through?+

A first working slice on your real documents in 3–6 weeks. Full rollout with integrations in 2–4 months.

What does it cost?+

Fixed-scope pilot first, priced from the saving we estimate together in hours per week. Most document projects pay back within 2–6 months.

Bring three real documents to the call.

Thirty minutes. We look at them together and tell you honestly what accuracy to expect and how many hours a week it frees.

Book a call with Artem