Recognica

Document Processing,
on your own server.

Recognica turns the PDFs your business receives every day into verified data. It reads them, recognises what each one is, captures the values you need, and has a person check only what deserves checking. It runs entirely inside your own network.

Unlimited documentsNo click chargesNo external AIOCR & text layers Machine-learning classificationField extractionHuman verification
Recognica logo

1 · Overview

From a pile of PDFs to clean, checked data

Invoices, claims, orders, forms: most organisations still re-key information from documents by hand. Recognica automates that work end to end. Documents are dropped into a shared folder (by people, scanners or other systems) and come out the other side as the original PDF plus a CSV file of the captured values, ready for your accounting, ERP or line-of-business system.

1

Reads any PDF

Uses the PDF's own text when it is there, and OCR when the page is a scan or photo, page by page.

2

Knows what it's looking at

A machine-learning model trained on your own examples sorts each document into its type, with a confidence score.

3

Captures the values

Document Definitions say where each field is found, such as invoice number, date, total or booth number.

4

People check the exceptions

Only documents and fields below your confidence thresholds go to a verifier, who sees the page with each value highlighted.

5

Delivers files you can use

Every finished document is exported as its PDF plus a CSV, in a folder per batch, on a Windows share.

6

Measures everything

Pages, documents, machine time and people's time are recorded for your reporting, along with how accurate the capture was.

Why Recognica

Who it is for

Companies that receive a steady flow of similar documents from many senders: accounts payable, logistics and freight, insurance claims, healthcare administration and back-office services, and service bureaus processing documents for their own clients. Customers, workflows and usage are kept separate, so one installation can serve several clients.

Pricing and privacy

Process as much as you want

Many document-capture products charge per page, per document or per "click", so your bill grows with your success. Recognica doesn't.

No click charges

No fees per page, per document or per field. Volume never changes the price.

Unlimited use

Unlimited importing, exporting, workflows and reprocessing, limited only by your server.

One monthly fee

A predictable monthly fee for support and professional services: set-up, new workflows, tuning and updates.

100% internal: no external AI

Nothing leaves your network

OCR, classification, field extraction, training and reporting all run on your own server. Documents, text and values are never sent to a cloud service.

No outside calls

Recognica makes no calls to external AI services or any other outside source while it works, so there's no third-party data processing to approve and no surprise usage bills.

Feature summary

2 · How a document gets in and out

The journey of a document

Each workflow has its own shared folders. The Conductor, Recognica's processing engine, watches them and moves every document through the same steps, timing each one.

Importshared folder Batchgrouped & logged Readtext layer / OCR Classifydocument type Extractfield values Verifyonly if needed ExportPDF + CSV Unknowna person picks Errorwith the reason low confidence type chosen unreadable / rejected sure enough

Documents arrive in Import

Anyone with access saves PDFs into the workflow's Import share: a person from Windows or a Mac, a network scanner, or another system. A lock file tells scanners to wait while Recognica is collecting files, so nothing is picked up half-written. System leftovers are cleaned up and stray subfolders are sent to Error with an explanation.

A batch is created

Everything collected together becomes a named batch (for example Invoices-20261113-0930-001), so it can be tracked, verified and reported as one unit.

Pages are read

For each page Recognica decides whether the PDF's own text layer is usable. If it is, it is used as is. If not, the page is rendered and read with OCR. A workflow can also force OCR, or use text layers only. The result is saved, so a document is never read twice.

The type is recognised

The classifier scores the document against every type it knows. Above the workflow's classification threshold, the type is set automatically. Below it, the document goes to the Unknown queue, where a person picks the type or rejects it, and can add it to the training examples.

Fields are captured

The Document Definition for that type finds each field's value and records its page, position and confidence.

A person verifies, only when needed

Documents with any field under the field threshold wait for verification. The verifier sees the page with each value boxed, corrects what's wrong, and confirms. Every correction is recorded with who made it and when.

Results are exported

Each finished document lands in the Export share as its original PDF plus a .csv of field names and values, in a folder per batch. Anything that could not be processed goes to Error, always with the reason.

2 · How a document gets in and out (continued)

Where people come in

Automation does the bulk of the work; people handle the judgement calls. Recognica gives them focused screens and keeps a full record of what they did.

Verification

  • Verifiers see the batches assigned to their workflows and reserve one, so two people never work on the same batch.
  • Each page is shown with every captured value boxed on it, with zoom and page navigation.
  • Fields below the confidence threshold are flagged check; the rest show their confidence.
  • Keyboard first: Enter moves to the next field and Ctrl+Enter confirms the document.
  • Release a batch for someone else, or cancel to undo everything changed in this sitting.

Unknown queue

  • Documents the classifier wasn't sure about wait here with the model's best guess and its confidence.
  • A person picks the right type (optionally verifying it right away) or rejects it to Error with a reason.
  • Picked documents can be added to the classification project as new training examples, so the model keeps learning.
  • Optionally, documents nobody decides on are rejected automatically after a set number of days.

Restarting and reprocessing

  • Send chosen documents, a batch, or a filtered set back to any step: OCR, classification, extraction, verification or export.
  • Reruns use the workflow's current published rules and model, and can keep values people already verified.
  • Each restart is marked billable or not billable, and the choice follows every timing it produces.

Housekeeping, automatically

  • Per-workflow retention for working files, Export, Error, and how long field values and detailed history are kept.
  • Usage and billing records are always kept, as are accuracy counts.
  • Nightly database backups, kept for 14 days.

What comes out

OutputWhereContents
Data fileExport share, folder per batchOne .csv per document: a header row of field names and one row of values, next to the PDF with the same name.
Original PDFExport shareThe document exactly as received, so the image and the data travel together into your systems.
Error filesError shareDocuments that couldn't be read, were rejected, or arrived in the wrong place, each with the reason in its history.
HistoryWorkflows app, ReportsEvery step, timing, decision and correction for every document, searchable by file, batch, type, status or person.
Integration made simple: because input and output are ordinary network folders and CSV files, Recognica works alongside almost any ERP, accounting package, RPA tool or scripted import, with no custom connector needed.

3 · Creating a workflow

Set up a new document process in an afternoon

A workflow is one document process for one customer, for example "Vendor invoices for Acme". Administrators build it in the browser with three tools that share one sign-in: Workflows, Document Classification and Document Field Extraction.

Add the customer Workflows

Every workflow belongs to a customer, so usage and billing are reported per customer and per workflow.

Teach Recognica your document types Document Classification

Create a classification project and drop example PDFs into a folder per type (invoice, credit note, statement…). Recognica reads them, builds the training corpus and trains a model with one click. Each training run is kept in the model history with its training log, so you can compare runs.

Describe the fields Document Field Extraction

Create a Document Definition per type and upload sample PDFs. For each field you add keyword rules and see the results live on every sample:

  • Keyword and direction: the value is to the right of, left of, below or above a label, or on the same line.
  • Variants: other wordings of the label, tried in order ("Invoice No", "Invoice #", "Inv. Number").
  • Value patterns: what a valid value looks like (date, amount, an ID format), so wrong neighbours are skipped.
  • Not words: words that cancel a match on the left or right ("ship" stops "Date" from matching "Ship date").
  • Rename, duplicate and run a whole definition to CSV to check it on all samples at once.

Create the workflow Workflows

Name it, pick the customer and the classification project, and choose the settings: classification and field confidence thresholds, how pages are read (automatic, force OCR, or text layer only), how often Import is checked, and retention for each kind of file.

Map document types to definitions

Tick the types this workflow accepts and choose the Document Definition for each. Types you don't tick are rejected to Error with the reason.

Choose people and folders

Assign verifiers to the workflow and create file-share accounts. Recognica creates the workflow's Import and Export shares, which PCs, Macs and scanners connect to like any Windows share.

Publish and start

Publishing freezes a copy of the settings, the rules for each type and the model. The Conductor runs that copy, so later edits never affect running batches. The workflow lists what changed since the last publication. Press Start, and drop the first PDFs into Import.

Features · Reporting and administration

Know exactly what was done, by whom, and how well

Overview

Documents and pages exported today, this week and in any period; what is waiting right now (in the machine steps, to verify, unknown, in Error); batches waiting for verification and who holds them; classification and field confidence per document type.

Usage

For your own reporting or for charging your clients back, per customer, workflow, day and person: documents read, exported and sent to Error; pages split by text layer and OCR; machine time per step; people's active time. Non-billable reruns are left out. Four CSV downloads that add up to the millisecond.

Verification accuracy

For every field, what was captured versus what the verifier confirmed: kept as captured, corrected, cleared, filled in or left empty, with accuracy as a percentage. A list of every correction shows the captured value, the verified value, who and when.

Timing

Average active time per step: per page for OCR, per document and per batch, by workflow. People's time stops after a minute without a key press or mouse movement, so figures reflect real work.

Search and document history

Find any document by file or batch name, type, status or person. Its history shows every event, every field (extracted value, current value, who corrected it), and every timing, with billable flags and reruns.

Server page

Live memory and CPU use per service, disk space and history, with limits that an administrator can change in the browser, including warnings when limits add up to more than the machine has.

Users, security and audit

Your data stays yours. Recognica does not send documents, text or values to any outside service. OCR, classification and extraction all run on your server, and processing needs no internet connection.

4 · Technology

Architecture

Recognica is a set of small, focused services that run in containers on one Linux server and talk to each other over a private internal network. Only the web entry point and the file shares are reachable from your network.

YOUR NETWORK Browsers: administrators & verifiers PCs, Macs & scanners (Windows shares) RECOGNICA SERVER · DOCKER COMPOSE Caddyweb entry · ports 80/8080/8081 File shareSamba · port 445 Workflows appsign-in · verify · reports Field ExtractionDocument Definitions Classificationprojects · training Conductorruns every workflow OCR enginePDFium · Tesseract Classifier engineONNX Runtime Trainerscikit-learn Serverhelpermetrics PostgreSQL 18workflows · documents · fields · billing · audit Document storageImport · Batches · Export · Error · models Backupsnightly
ServiceRole
CaddySingle web entry point. Routes to the three web apps and checks with the Workflows app that a person is signed in as an administrator before opening a design tool. Can serve HTTPS with automatic certificates.
Workflows appSign-in, customers, users, workflows and publishing, file-share accounts, verification, the Unknown queue, restarts, reports and the server page.
ConductorThe processing engine. Watches every running workflow's Import, creates batches, calls the engines in order, records timings and events, exports results and applies retention.
OCR engineReads text layers, renders pages and runs OCR, page by page in parallel, and times each page for billing.
Classifier engine & trainerThe engine scores documents with a project's model, or a workflow's published copy. The trainer builds new models from the training corpus.
Database, file share, helpersPostgreSQL holds all records. Samba serves the Import, Export and Error shares. A migration job applies database updates and a backup job dumps the database nightly. The server helper reports resource use and is the only service allowed to adjust container limits.

4 · Technology in detail

The technology inside

Reading documents

Text layerPdfPig (.NET) reads the text and word positions already inside digital PDFs. A usability check (characters per page, share of printable characters) sends weak or garbage text layers to OCR instead.
RenderingPDFium, the PDF engine used in Google Chrome (via Docnet), renders scanned pages to images at the resolution OCR needs.
OCRTesseract 5, the open-source OCR engine originally developed at HP and later Google, using its neural-network (LSTM) recogniser. It returns every word with its position and confidence. Pages of a document are read in parallel.
Saved resultsEach page's text, words, positions and confidence are saved next to the PDF, so later steps, restarts and the design tools never have to read the same page again.

Recognising the document type

ModelA text classifier: TF-IDF features over words and word pairs (bigrams) feeding a class-balanced logistic regression. It is fast, explainable, and accurate on business documents with a modest number of examples per type, with no GPU needed.
TrainingBuilt with scikit-learn (Python 3.12) from the text of your labelled examples, with pinned library versions so results are reproducible. Each run is logged and kept in the model history.
RunningModels are exported to ONNX, an open standard for machine-learning models, with the vocabulary built in, and run by Microsoft ONNX Runtime inside the .NET classifier engine. The same normalised text is used for training and prediction.
ConfidenceEvery prediction carries a probability per type. The workflow's threshold decides between automatic filing and the Unknown queue.
No external AI. Every model is trained and run on your server. Recognica sends nothing to cloud AI services and makes no calls to any outside source, so there are no per-call fees, no data sharing and no dependency on someone else's service.

4 · Technology in detail (continued)

Extraction and platform

Capturing fields

Rules engineRecognica's own keyword-based extraction engine works on the positioned words of each page. It handles labels and values on the same line or across lines, in four directions, with label variants, value patterns (regular expressions) and Not-words to the left or right.
ConfidenceEach value carries a confidence from the OCR of its words, compared with the workflow's field threshold to decide whether a person should check it.
VersioningRules are saved as JSON with a checksum. Publications store a frozen copy, and every field records which rule version produced it, so reruns can keep verified values where the rule didn't change.

Platform

ApplicationsC# on .NET 8 (Microsoft's long-term-support release). The web apps use ASP.NET Core Blazor Server for responsive screens without browser plug-ins. Data access uses Npgsql and Dapper.
DatabasePostgreSQL 18, with versioned migrations applied automatically on start-up, column-level permissions per service, and triggers that keep billing and audit tables append-only.
Web & filesCaddy 2 web server and reverse proxy. Samba for SMB file sharing, compatible with Windows, macOS, Linux and network scanners.
RuntimeDocker containers orchestrated with Docker Compose on Linux. Each service has its own memory and CPU limits, health checks and automatic restart. It runs on both ARM64 and x86-64 processors.

Built to be dependable

4 · Servers and deployment

Runs on hardware you choose

Recognica is delivered as one stack of containers for a single Linux server: an on-premises machine, a virtual machine in your data centre, or a compact edge device. The same software runs on ARM64 and x86-64.

ServerTypical useSpecification
Edge applianceDemonstrations, pilots, a single department with light volumeNVIDIA Jetson Orin Nano (ARM64, 6 cores, 8 GB), NVMe storage, Ubuntu-based JetPack Linux. Recognica's reference development system; fits in the palm of a hand.
Standard server or VMMost production installationsx86-64 Linux, e.g. Ubuntu Server 24.04 LTS · 4+ CPU cores · 16 GB RAM · 100 GB disk plus document storage · fixed IP address.
RecommendedHigher volumes, several customers or workflows, frequent retraining8+ CPU cores · 32 GB RAM · SSD storage sized to your retention settings. OCR, the busiest step, scales with CPU cores.

No GPU is required: OCR and classification run on the CPU. Memory and CPU limits for each service are set to fit the machine and can be changed from the Server page.

Network

  • Web: port 80 (Workflows), 8080 (Field Extraction) and 8081 (Classification), or HTTPS with your own host name.
  • File shares: SMB on port 445.
  • Everything else stays on the server's private internal network.
  • No internet connection, and no outside AI service, is needed to process documents.

Data and backups

  • Documents, models and the database are kept in one data folder on the server.
  • The database is dumped every night and kept for 14 days, with a ready-made script to copy dumps to another machine.
  • Retention settings keep storage use predictable.

Installation

  • A single command starts the whole stack; database updates apply themselves.
  • A one-time setup code creates the first administrator in the browser.
  • A self-contained offline installation package for new servers is in development.

Updates and versions

  • Every page shows the running version.
  • Workflows keep running on their published version while the software or rules are updated.
  • Source and releases are kept in Recognica's own private version control.

Let's look at your documents.

The fastest way to see what Recognica can do is with your own paperwork. Send us a handful of sample PDFs and we'll show you them classified, captured and exported, on a server that never leaves your control.

Unlimited documents · No click charges · No external AI · One monthly fee

[email protected]  ·  813-830-1708

Product names mentioned (NVIDIA, Jetson, Microsoft, .NET, PostgreSQL, Tesseract, Docker and others) are trademarks of their respective owners.