Solutions / Private IDP (Intelligent Document Processing) — Intelligent Document Processing and Verification for Business
Private IDP is an enterprise system that ingests an applicant's document package: scans, phone photos, PDFs, and Word files mixed together. The system recognizes the package, identifies the type of each document, extracts key fields, and verifies the package against your organization's requirements.
Processing runs entirely on your server. Documents, their contents, and extracted data are never transmitted to cloud services or external APIs.
The Challenge
An applicant submits a document package in various formats. A single PDF may contain six different documents, while a single document may be split across three photos. An employee has to manually sort out what each file is, what is missing, which certificates have expired, and whether the data matches across documents.
Processing a single package takes hours. Errors are caught at the next stage, when bringing the applicant back for a missing document is already costly.
What the System Does
Private IDP performs incoming package validation in minutes:
- recognizes every page — printed text, tables, and details, including on scans and photos;
- classifies pages by document types from your directory and assembles them into complete documents;
- extracts required fields — IIN/BIN, dates, amounts, numbers, and entity names;
- verifies the package — whether all required documents are attached, mandatory fields are filled out, validity periods have not expired, and data reconciles across documents.
The output is an actionable report with specifics: which document is missing and under which requirement it is mandatory, which field is blank, and which page is illegible and needs to be re-photographed. Each extracted value links directly to the source page and snippet — verifiable in a single click.
The final decision remains with your employee: the system prepares the groundwork rather than issuing a verdict.
Privacy as a Core Architectural Principle
Applicants' documents contain personal data, financial information, and—for banks—information constituting banking secrecy. That is why the system was engineered from the ground up so that data never leaves your environment.
- Dedicated installation for each client. The system is deployed on your server—in your server room or in your own data center rack. There is no shared cloud or shared database with other clients.
- All models run locally. Recognition, classification, and field extraction are performed by open-source models on your hardware. Files, text, and extracted values are never sent to model developers or to us.
- Operation in an air-gapped network segment. Once installed, the system requires no internet access.
- No document archive stored by the system. Files remain in your core system or EDMS; the service receives them only for the duration of processing. Intermediate files are deleted immediately after processing, and results are purged according to a configurable retention period.
- Data localization. Data never leaves the client's infrastructure or territory, simplifying compliance with personal data storage and banking secrecy regulations.
- Auditability. The raw recognition output is preserved unchanged. Every system conclusion can be traced back to the exact page of the source document.
How It Works
1. File-Type Routing
Each format is processed using the optimal method. Text is extracted directly from Word files without OCR. For PDFs, the text layer is checked first: if present, the existing text with coordinates is extracted; if it is a scan, the page is rasterized at 200–300 dpi and sent for recognition. This is faster and more accurate than running OCR on everything: in real-world packages, a significant share of pages does not require OCR at all.
2. Photo Preprocessing
For phone photos, the system detects page boundaries, corrects perspective and rotation, and normalizes lighting. Page quality is assessed prior to recognition: if an image is unusable, the system immediately requests a retake of that specific page rather than flagging the issue the next day.
3. Optical Character Recognition (OCR)
Recognition is powered by compact, specialized document models of around 1 billion parameters. On open document parsing benchmarks, these models outperform general-purpose multimodal neural networks that are hundreds of times larger, while running on a single server. The output includes text with bounding-box coordinates, tables, and stamp detections. The system is built for Russian and Kazakh, including the specific letters of the Kazakh Cyrillic alphabet.
4. Document Classification and Assembly
Each page is assigned to one of the types in a closed document directory, after which pages are assembled into documents. Multiple documents in a single file and a single document split across multiple files are handled identically.
5. Schema-Based Field Extraction
Each document type has its own defined set of fields. If a value is illegible or missing, the system returns "not found" rather than fabricating a plausible guess. Fields with low recognition confidence are flagged for manual review.
6. Rule-Based Verification, Not a Neural Network
Completeness and compliance with requirements are checked by a deterministic rule engine. The output is reproducible: the same package always produces the exact same report, and every finding references a specific rule. Requirements are configured via settings and can be updated by a business analyst without developer involvement.
7. Cross-Document Checks
Matching IIN/BIN across all documents in the package, checking certificate expiration dates, and reconciling amounts between financial statements and the application.
Use Cases
The primary industry scenario is loan application package intake. The exact same task—ingesting, parsing, and validating a document package against requirements—arises in other workflows as well:
- corporate client onboarding and KYC;
- tender and procurement bids;
- HR onboarding documents;
- insurance and leasing applications.
Integration and Infrastructure
- REST API with token authentication. Jobs run asynchronously: your system submits a package and retrieves the result once ready.
- Output is delivered in two formats: structured data (JSON) for your system and a human-readable report for the employee.
- The system runs alongside your backend or EDMS rather than replacing them.
- Built on the same platform as Private ASR: the job queue, unified API, logs, and monitoring are already proven in production.
Local Server Requirements
A single server with a GPU featuring at least 16 GB of VRAM is sufficient: the document recognition and text analysis models reside in memory simultaneously without model swapping. A 100-page package is processed in minutes.
See Also
Private ASR — enterprise speech recognition system deployed within your company's infrastructure.






