AI Form Autofill from Documents: Donut, LLM & PyTorch

Note: when a document is scanned, standard OCR outputs flat text without structure. An address might end up in the "Full Name" field, and a date in the notes. This situation is typical in accounting, where manual data entry takes hours. We solve this problem with a Document Understanding pipeline — it not only recognizes characters but understands where each piece belongs. In practice, this speeds up data entry 5–10 times and reduces errors by 70%. For a team of five operators, savings can reach 1.5 million rubles per year. The system pays for itself in 3–6 months through reduced labor costs.

Problems We Solve

• Unstructured text after OCR. Classic Tesseract or Google Vision produce raw text with coordinates but don't know that "Ivanov" is a last name and "123456" is a passport number. We use LayoutLM v3, which analyzes both text and block geometry.
• Different document templates. An invoice from one supplier differs from another, and a passport changes page layouts. Our pipeline classifies the document type and applies a specialized extractor.
• Low accuracy on complex documents. Invoices with tables, handwritten notes, low-quality scans — all break regular OCR. We add an LLM with vision that understands context and recovers missing information.

How AI Autofill from Documents Speeds Up Data Entry?

The core is a three-stage pipeline: classification → extraction → mapping. For standard types (passport, TIN) we use LayoutLM trained on thousands of labeled samples. For non-standard ones — multimodal LLM that processes the image and returns JSON with fields.

from typing import Any
from dataclasses import dataclass

@dataclass
class FormField:
    name: str           # имя поля в целевой форме
    value: Any          # извлечённое значение
    confidence: float   # уверенность 0..1
    source_location: str | None = None  # откуда извлечено

class DocumentToFormPipeline:
    """
    Пайплайн: документ → поля формы.
    Шаги:
    1. Классификация типа документа
    2. OCR + layout analysis
    3. NER/IE для извлечения полей
    4. Маппинг на поля целевой формы
    5. Валидация и нормализация
    """
    def __init__(
        self,
        document_classifier,   # модель классификации типа документа
        extractors: dict,      # {doc_type: extractor}
        form_mapper: dict,     # {doc_field: form_field} маппинг
        validators: dict       # {form_field: validator_fn}
    ):
        self.classifier = document_classifier
        self.extractors = extractors
        self.form_mapper = form_mapper
        self.validators = validators

    def process(self, document_image) -> dict[str, FormField]:
        # Шаг 1: определяем тип документа
        doc_type = self.classifier.predict(document_image)

        if doc_type not in self.extractors:
            raise ValueError(f'Unsupported document type: {doc_type}')

        # Шаг 2-3: извлечение полей
        extractor = self.extractors[doc_type]
        raw_fields = extractor.extract(document_image)

        # Шаг 4: маппинг
        form_fields = {}
        mapping = self.form_mapper.get(doc_type, {})

        for doc_field, value in raw_fields.items():
            if doc_field in mapping:
                form_field_name = mapping[doc_field]
                form_fields[form_field_name] = FormField(
                    name=form_field_name,
                    value=value.get('text'),
                    confidence=value.get('confidence', 0.0),
                    source_location=doc_field
                )

        # Шаг 5: валидация
        for field_name, field in form_fields.items():
            if field_name in self.validators:
                try:
                    field.value = self.validators[field_name](field.value)
                except Exception as e:
                    field.confidence *= 0.5   # снижаем уверенность при ошибке валидации

        return form_fields

Hybrid LayoutLM and LLM — Advantage Over Classic OCR

Classic OCR doesn't understand semantics: the word "Moscow" could be a city or a street name. LLM with vision (e.g., Claude 3.5 or GPT-4o) analyzes the entire document at once: it sees field locations, related headers, and tables. This boosts accuracy by 15–20% on complex layouts. We use LLM as a fallback for rare document types. OCR alone does not solve structure understanding.

import anthropic
import base64
from pathlib import Path

def extract_form_fields_llm(
    document_path: str,
    form_schema: dict,    # JSON Schema целевой формы
    model: str = 'claude-opus-4-5'
) -> dict:
    """
    Мультимодальный LLM как универсальный Document Understanding engine.
    form_schema: {'field_name': {'type': 'str', 'description': '...', 'required': bool}}
    """
    client = anthropic.Anthropic()

    # Загружаем документ как base64
    with open(document_path, 'rb') as f:
        doc_b64 = base64.standard_b64encode(f.read()).decode()

    ext = Path(document_path).suffix.lower()
    media_type = {
        '.jpg': 'image/jpeg', '.jpeg': 'image/jpeg',
        '.png': 'image/png',  '.pdf': 'application/pdf'
    }.get(ext, 'image/jpeg')

    # Формируем описание схемы формы
    schema_desc = '\n'.join([
        f'- {name}: {info["description"]} ({"обязательное" if info.get("required") else "опциональное"})'
        for name, info in form_schema.items()
    ])

    message = client.messages.create(
        model=model,
        max_tokens=2048,
        messages=[{
            'role': 'user',
            'content': [
                {
                    'type': 'image',
                    'source': {
                        'type': 'base64',
                        'media_type': media_type,
                        'data': doc_b64
                    }
                },
                {
                    'type': 'text',
                    'text': f"""Извлеки из документа следующие поля для заполнения формы:

{schema_desc}

Верни результат строго в формате JSON:
{{
  "field_name": {{"value": "...", "confidence": 0.0-1.0, "not_found": false}},
  ...
}}

Если поле не найдено в документе, укажи "not_found": true.
Для confidence: 1.0 = точно уверен, 0.5 = сомневаюсь."""
                }
            ]
        }]
    )

    import json
    try:
        return json.loads(message.content[0].text)
    except json.JSONDecodeError:
        # Извлекаем JSON из текстового ответа
        import re
        match = re.search(r'\{.*\}', message.content[0].text, re.DOTALL)
        return json.loads(match.group()) if match else {}

How We Ensure Extraction Accuracy?

We combine three verification levels:

• LayoutLM outputs confidence for each field based on trained model.
• LLM additionally checks logical consistency (e.g., birth date not later than today).
• User-in-the-loop: fields with confidence < 0.85 are highlighted in the UI for manual review. This is standard practice in enterprise solutions.

def prepare_form_ui_state(
    form_fields: dict,
    confidence_threshold: float = 0.85
) -> dict:
    """
    Подготовка состояния формы для UI:
    - Поля с высокой уверенностью — автозаполнены
    - Поля с низкой — помечены для проверки
    - Обязательные не найденные — ошибка
    """
    ui_state = {}
    for field_name, field in form_fields.items():
        status = 'autofilled'
        if field.value is None:
            status = 'not_found'
        elif field.confidence < confidence_threshold:
            status = 'needs_review'

        ui_state[field_name] = {
            'value': field.value,
            'status': status,
            'confidence': field.confidence,
            'editable': status != 'autofilled' or True  # всегда редактируемо
        }
    return ui_state

Approach Comparison Table

Timelines

What's Included

• API documentation (Swagger, Postman collection) and pipeline description.
• Model training on your samples (up to 100 documents free).
• Trial period — 2 weeks with engineer support.
• Integration — ready modules for 1C and Bitrix24.
• Maintenance — 95% accuracy guarantee for 3 months, SLA on bug fixes.

Our team has 10 years of experience in Computer Vision and NLP and has delivered over 50 Document Understanding projects. We guarantee the system fits your landscape without delays. Request a demo version for testing on your documents. Contact us for a project evaluation — we'll prepare a prototype in 2 weeks.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.