Developing Document AI for Document Processing

Your operators spend hours entering invoices and waybills, yet errors still slip through? We've encountered this with dozens of clients and know how to fix it. Processing 1,000 invoices takes up to 80 operator hours, costing thousands of dollars monthly. In our practice, we've built dozens of intelligent document processing (IDP) solutions for banks, logistics, and retail. Let me explain how such systems are built and what results can be achieved.

Document AI processes documents tens of times faster than a human, and cost savings per document reach 95% compared to manual entry. For a client with a volume of 2,000 documents per month, this means savings of 2 to 5 million rubles annually. Let's compare manual processing vs. Document AI:

How Document AI Saves Up to 95% of Processing Time

The platform consists of several layers, each solving its own subtask. A typical pipeline:

[Incoming document: PDF, DOCX, JPG, TXT]
    → [Document Intake Service]
         ├── Type detection (PDF/scan/text)
         └── Routing

    → [Pre-processing]
         ├── OCR (if scan): Tesseract / Azure DI / Google Document AI
         ├── Table extraction: Camelot / pdfplumber / Table Transformer
         └── Layout Analysis: page element positioning

    → [AI Processing Pipeline]
         ├── Document type classification
         ├── Structured data extraction
         ├── Extracted data validation
         └── Summarization / analytics generation

    → [Output]
         ├── JSON with extracted data
         ├── Structured database record
         └── Notifications / integrations

The process of building such a system includes five stages:

1. Collect and label a representative sample of documents (100–200 pieces).
2. Select and configure OCR + image preprocessing.
3. Train classification and field extraction models.
4. Integrate with ERP via REST API or direct connectors.
5. Set up monitoring and a retraining loop as new types arrive.

Why Is Preprocessing a Critical Stage?

OCR and layout analysis quality directly determines the accuracy of subsequent extraction. For low-resolution scans, convolutional networks (Table Transformer) must be applied to detect tables. In one retail project, we dealt with waybills printed on thermal paper—we had to retrain a model on synthetic data simulating fading. Using modern OCR engines like Azure Document Intelligence improves extraction accuracy to 99% on quality scans, but custom preprocessing is required for complex cases.

Document Types Processed

The system handles various document types. Approaches differ:

• Structured (invoices, waybills, tax forms): deterministic formats, high-accuracy field extraction.
• Semi-structured (contracts, forms, applications): variable structure, requires context understanding.
• Unstructured (letters, reports, medical records): free text, NLP processing.
• Images and scans: preliminary OCR, then NLP processing.

Technical Implementation

Data Extraction from Structured Documents

For standardized forms (invoices, UPD, waybills in Federal Tax Service XML format) — deterministic parsing via XPath, no ML:

from lxml import etree

def parse_upd(xml_path: str) -> InvoiceData:
    tree = etree.parse(xml_path)
    root = tree.getroot()
    ns = {"n": "urn:NDS"}

    return InvoiceData(
        seller_inn=root.findtext(".//n:СвПродавца/n:ИдСв/n:СвЮЛ/@ИННЮЛ", namespaces=ns),
        invoice_number=root.findtext(".//n:Документ/n:НомерДок", namespaces=ns),
        total_amount=float(root.findtext(".//n:ВсегоОпл", namespaces=ns) or 0),
    )

ML is only needed for non-standard formats.

IDP for Scans: Stack and Examples

from azure.ai.documentintelligence import DocumentIntelligenceClient
from azure.core.credentials import AzureKeyCredential

client = DocumentIntelligenceClient(endpoint, AzureKeyCredential(key))

# Analyze an invoice
with open("invoice.jpg", "rb") as f:
    poller = client.begin_analyze_document(
        model_id="prebuilt-invoice",
        body=f.read(),
        content_type="application/octet-stream"
    )
result = poller.result()

# Access fields
invoice = result.documents[0]
vendor_name = invoice.fields.get("VendorName")
total_amount = invoice.fields.get("InvoiceTotal")

Alternatives: Google Document AI, AWS Textract, PaddleOCR + LLM extraction for on-premise. Learn more about OCR.

Classification and Data Validation

Multi-class classifier based on:

• Text content (TF-IDF / BERT embeddings)
• Structural features (presence of tables, number of pages, sections)
• Metadata (file name, source)

Typical accuracy: 96–99% for distinct types (invoice vs contract vs act), 88–94% for similar types.

Each extracted field is accompanied by a confidence score. At low confidence (< 0.8) — flagged for manual review. Cross-validation: do amounts in words and digits match? Does TIN pass checksum? Is date logical? Straight-Through Processing rate for high-quality structured documents reaches 85–95%. Time and cost savings: automation ROI in 6–12 months for medium businesses. Many clients recoup the investment in 8–10 months.

Result and Pilot

What's Included in the Final Solution

• API with endpoints for upload, processing, and data retrieval
• OCR module supporting Tesseract, Azure DI, or Google DI
• Document type classifier (fine-tunable model)
• Field extraction with confidence scores
• Validation and logging
• Web interface for manual correction and monitoring
• Integration with 1C, SAP, or other ERP (REST/SOAP/file exchange)
• Documentation and operation manual
• Operator training (2 days)
• Pipeline warranty for 6 months

Implementation Timeline

Month 1: OCR pipeline, document type classifier, basic extraction

Months 2–3: Processing priority document types, validation, ERP/ECM integrations

Month 4: Semantic search, manual correction UI, analytics

Months 5–6: Production hardening, scaling, quality monitoring

How to Start a Pilot?

A pilot project begins with collecting a sample of 100–200 typical documents (scans, PDFs, XML). In 1–2 days, we evaluate architecture and extraction accuracy, and prepare a preliminary cost estimate. After agreement, the full pipeline is launched—from OCR to integration. First results are visible in 2–3 months. Then we refine all document types, connect ERP, and train operators.

Order a pilot project to assess effectiveness on your documents. Contact us — we'll prepare a preliminary architecture in 1–2 days and evaluate the project. Get a consultation on integrating Document AI into your infrastructure.

NLP Development: Text Classification, NER, Embeddings, and Information Extraction

We often receive a task: process 50,000 support tickets — currently all manual. Dataset — 3,000 labeled examples, 12 categories, imbalance: one category occupies 40% of the sample, three at 1-2% each. Baseline accuracy — 78%. Sounds decent until you look at recall for rare classes: 0.31, 0.44, 0.28. These classes — complaints and churn threats — are most important to the business.

This is a typical NLP development project. The problem is not the algorithm but that accuracy is the wrong metric. Our experience across 30+ projects shows: we start by analyzing business metrics and only then choose the model.

Why accuracy is not the right metric for rare classes?

Accuracy ignores imbalance. If the "churn" class appears in 2% of cases, the model can predict "all good" and get 98% accuracy — but the business loses clients. Solution: F1 macro (averaged over all classes) or weighted F1. For NER — strict entity F1 (exact matches only). We guarantee: after choosing the correct metric, model quality becomes measurable and predictable.

Text Classification: From BERT to Distillation

BERT-like models are the standard for classification. ruBERT-base or ruBERT-large from DeepPavlov for Russian. multilingual-e5-large — for multiple languages in one pipeline. XLM-RoBERTa-large — a strong multilingual backbone.

Fine-tuning for classification: add a classification head on top of the [CLS] token, train for 3-5 epochs with lr=2e-5, weight decay=0.01. For imbalance — weighted CrossEntropyLoss or focal loss with gamma=2.0. Contact us — we will show a code snippet.

Imbalance case study. Dataset — 3,000 examples, imbalance 1:20. Solution: class_weight via sklearn + CrossEntropyLoss. Additionally — augmentation of rare classes via backtranslation (ru→en→ru through MarianMT). Recall for rare classes rose from 0.31 to 0.67 with a slight drop in accuracy (76%→74%). Full NLP development end-to-end took 3 weeks.

Distillation for production. BERT-large gives F1 0.89, but inference on CPU — 180ms. Distillation into DistilBERT or ruBERT-tiny2 reduces latency to 25ms with F1 0.84. Export to ONNX Runtime provides an additional 1.5-2x speedup. DistilBERT achieves 7x lower latency than BERT-large with only a 5% drop in macro F1 – a typical production trade-off.

How to choose between BERT and LLM for your task?

For most classification and extraction tasks, BERT-sized models offer the best trade-off between cost and performance. Shift to LLMs only when the task demands generation, complex reasoning, or zero-shot generalization.

NER: Named Entity Recognition

NER — extracting persons, organizations, locations, dates, amounts, document numbers. For general categories (PER, ORG, LOC), pre-trained models work well. For specialized ones (medical terms, legal concepts) — fine-tuning is needed.

Data annotation. The main cost of an NER project. For a quality model — 500-2,000 labeled sentences per entity type. Tools: Label Studio (open source) or Prodigy (by spaCy creators). IOB2 format — standard.

Architecture. Token classification on top of BERT: each token gets a label (B-PER, I-PER, O). spaCy 3.x with transformer pipeline — a convenient production choice.

Nested entities. Standard IOB models cannot handle nested entities (organization inside an address). For such tasks — span-based NER: SpanBERT or SpERT. More complex but correct.

Post-processing is mandatory. The model predicts tokens — normalized entities are needed. Date — dateparser. Amounts — regex + validation. Names — deduplication via rapidfuzz. Included in our standard delivery.

Sentiment Analysis and Opinion Mining

Binary classification positive/negative works out of the box with BERT. Complexity — aspect-based sentiment analysis (ABSA): "the restaurant has good food but terrible service." For ABSA: aspect extraction (NER) + sentiment per aspect. Joint models BERT-for-ABSA — quality on Russian data is lower due to dataset scarcity. RuSentiment, SentiRuEval — main resources.

For production with simple positive/negative/neutral: distil models are enough. Three classes, balanced dataset, 2,000+ examples — F1 macro 0.82-0.87 in 1-2 days.

Text Summarization

Extractive summarization (select sentences) — TextRank or BM25 without training. Fast, no hallucinations. Good for long documents.

Abstractive (generates new text) — seq2seq: mT5, mBART, FRED-T5, ruT5-large. For production via LLM API (GPT-4, Claude) — often the best cost/quality/speed trade-off.

Embeddings: Vector Representations of Text

Embeddings are the foundation of semantic search, deduplication, clustering, RAG. Quality critically affects downstream tasks.

Models. E5-large-v2, BGE-M3, multilingual-e5-large — strong multilingual embedders. sentence-transformers/paraphrase-multilingual-mpnet-base-v2 — fast option. For Russian: ru-en-RoSBERTa (Skoltech) performs well on semantic textual similarity.

Embedding quality evaluation uses the MTEB benchmark as standard. But top results on MTEB don't guarantee success on a domain dataset — we build domain-specific eval.

Fine-tuning embeddings. If standard models don't give the required Recall@k — contrastive learning on domain pairs with MultipleNegativesRankingLoss. How to perform this for domain data:

1. Collect 500–2,000 semantically similar pairs from your domain.
2. Apply MultipleNegativesRankingLoss with a batch size of 32–64.
3. Train for 1–3 epochs using AdamW (lr=2e-5).
4. Evaluate Recall@k on a held-out domain test set.

This approach yields a 5–15% improvement in Recall@k in practice.

Dimensionality and storage. E5-large: 1024 dim, float32 — 4KB per vector. For 10M documents — 40GB. Quantization int8 reduces to 10GB. FAISS IVF_PQ — more compact but with losses. Included in our deployment recommendations.

Information Extraction

Structured extraction is a frequent task. Examples: key contract terms, technical characteristics, dates and amounts from invoices.

1. Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
2. NER + post-processing. For variable formats.
3. LLM with structured output. GPT‑4 / Claude with JSON schema — for complex documents. Cost: minimal per document. For 10k+ documents/day — we calculate the economics.

We guarantee a hybrid: regex/NER for typical fields + LLM for edge cases. Our guarantee is backed by years of production experience and more than 30 projects.

Work Stages

Prototype on existing data — 1-3 weeks. Production system with CI/CD — 1.5–2.5 months. Cost is calculated individually — get a consultation for a project estimate.

What's Included

• Model and pipeline architecture documentation
• Access to the model via REST API (FastAPI + ONNX)
• Client team training (2-hour webinar + Q&A)
• Accuracy guarantee on the agreed test set
• Months of post-delivery support (bug fixes, adaptation to new data)

Our Experience

Years of NLP projects from classification to RAG systems. The team includes ML engineers experienced with Hugging Face, spaCy, LangChain, MLOps. We use vLLM, Kubeflow, Weights & Biases — a production stack, not toys. Contact us to evaluate your NLP project within two days — request a free consultation on your text processing pipeline.