Document AI, speech-to-text, and domain-specific models

Make documents and recordings ready for AI.

Extract structured information from complex documents, transcribe recordings by speaker, and apply language models tuned to your organization’s terminology and tasks.

Talk to our team Explore demos

From scanned pages and recordings
to information your teams can use.

DocuLLM combines document recognition, speech-to-text, and domain-specific models to support search, analysis, and document-based workflows.

AI OCR

Optical character recognition (OCR) that uses Fast OCR for high-volume documents and VLM OCR for complex ones, keeping tables and lists intact.

Speech-to-text

Speech-to-text (STT) with speaker diarization turns recordings and meeting audio into searchable, speaker-labeled transcripts.

Domain-specific models

Fine-tuned on your terminology, criteria, and writing style so the output fits the task at hand.

Multi-model routing

Send each task to a security- or law-specialized model or to a general-purpose LLM, depending on the work.

Read, transcribe, and adapt

AI OCR
Extract more than plain text.
Use Fast OCR for standard documents and VLM OCR for complex layouts. Capture headings, tables, and key fields so the output can support search, summarization, and Q&A.
- Process HWP/HWPX, PDF, Office documents, scans, and images
- Recognize tables, multi-column layouts, and handwriting
- Extract fields such as dates and amounts
- Select the OCR engine according to document characteristics
Speech-to-text
Turn recordings into searchable,
speaker-labeled transcripts.
Speech-to-text with speaker diarization turns audio and video into transcripts.
Metadata makes interviews, consultations, and meetings easy to search and review.
- Transcribe audio from recordings and video files
- Separate transcript segments by speaker
- Retain speaker and segment metadata
- Support meetings, interviews, consultations, and investigation records
Domain-specific models
Adapt models to the way
your organization works.
Fine-tune models using domain materials such as past decisions, reports, Q&A, and regulations. Reflect the terminology, classification criteria, and writing patterns your tasks require. Route requests between domain-specific and general-purpose models to match the needs of each task.
- Domain-specific LLM and SLM development
- Fine-tuning and alignment
- Task-specific evaluation and model serving
- Multi-model routing and orchestration

Prepare the data.
Apply the protection it needs.

Connect DocuLLM with DocuSafety to detect and de-identify sensitive information in OCR and speech-to-text output. Apply input and response controls as that information is used with AI.

DocuSafety integration

De-identify personal and sensitive information

Detect names, contact details, national ID numbers, and other sensitive information in OCR and speech-to-text output. Apply masking, substitution, or pseudonymization according to policy before the data is used for search, summarization, and Q&A.

Detection of PII (personally identifiable information) in OCR output and transcripts
Masking, substitution, and pseudonymization
Policy-based rules and processing history

Multi-modal Guardrails Engine integration

AI DLP

AI data loss prevention (DLP) inspects model inputs and responses to help prevent leaks of personal and sensitive information, and blocks prompt injection and harmful requests. Attached documents and images are checked along with text, whether you use general-purpose or domain-specific models.

Sensitive data blocking at the input and response stages
Prompt injection and harmful request detection
Inspection of attached documents and images

PII

Personal data detection and de-identification

OWASP

Security controls reviewed against the OWASP Top 10 for LLM Applications

Explore DocuSafety →

Run document AI in the environment
your work requires.
Deploy DocuLLM through APIs or containers in on-premises, private cloud, or hybrid environments. Use locally deployed models where external model access is restricted.
As the document and model engine for DocuRAG, DocuLLM handles document and speech recognition and model inference for DocuPortal chatbots and DocuFlow agents and workflows.
Start with DocuLLM, then connect additional DocuRAG modules for chatbots, workflows, and data protection.

Who it’s for

Turn scans and recordings into searchable data

Convert petitions, receipts, inspection checklists, and contracts stored as scans, photos, or HWP files, along with investigation, consultation, and meeting recordings, into data your teams can search and analyze.

Bring your own criteria into AI work

Reviews, appraisals, legal analysis, and National Assembly responses rely on specific terminology and criteria. Domain-specific models help where a general-purpose LLM alone struggles to meet the required quality.

Run AI where external models are restricted

Public institutions and enterprises on closed or internal networks can build, serve, and operate their own models for each task.

What changes with DocuLLM

01
Scans, images, and speaker-labeled recordings converted into data AI can use
02
OCR matched to each document, balancing accuracy and processing cost
03
Less time spent retyping document contents
04
Output that reflects your organization’s terminology and criteria
05
The flexibility to use the right model for each task

Frequently asked questions

What document formats can it read?

DocuLLM reads HWP/HWPX, PDF, and Office files, as well as scans, photos, and screenshots. Fast OCR handles high-volume standard documents, and VLM OCR handles complex layouts such as tables, multi-column pages, handwriting, and low-quality images.

Does it handle recordings and meeting audio?

Yes. DocuLLM extracts audio from audio and video files and transcribes it with speaker diarization, producing transcripts organized by speaker with segment metadata. It can be applied to investigation recordings, consultations, meetings, and call center calls.

How is it different from a general-purpose LLM?

DocuLLM uses models fine-tuned on your organization’s terminology, criteria, and writing style, so output better fits your tasks. We develop domain-specific LLMs and small language models (SLMs) through fine-tuning and alignment, and route requests between specialized and general-purpose models as each task requires. In internal task evaluations, a CoT (chain-of-thought) fine-tuned domain model reached 96% search and reasoning accuracy; results vary by domain and data.

Can it run on a closed network where external LLMs aren’t allowed?

Yes. DocuLLM can be deployed through APIs or containers in on-premises, private cloud, or hybrid environments. Where external model access is restricted, it runs on locally deployed models.

Can we start with DocuLLM on its own?

Yes. DocuLLM serves as the document and model engine for DocuRAG, so you can start with it and connect additional DocuRAG modules for chatbots, AI agents, workflows, and data protection later.

How is personal information in OCR and speech-to-text output handled?

When DocuLLM is connected to DocuSafety, personal and sensitive information in OCR and speech-to-text output, such as names, contact details, and national ID numbers, is detected and masked, substituted, or pseudonymized according to policy. The processed data can then be used for search, summarization, and Q&A, subject to your access and usage policies.

Build on the documents and recordings you already have

Discuss your source formats, target workflows, and deployment requirements with our team.