Korean National Police Agency: An investigation support AI platform that makes case records usable by AI
Background
Policing and investigation work involves records in a wide variety of forms. Complaints and accusations, written statements, and supporting documents often arrive as scanned images or paper, and interviews also produce audio recordings. Records in such mixed formats take a great deal of time to search and organize, and as they are, they are hard to use for search or AI analysis.
The security requirements are strict as well. Investigation records contain a great deal of personal and sensitive information, such as names, phone numbers, and addresses, so protective measures are essential before any AI use. The system also had to run on an internal network separated from the outside, and processing history had to be available for later review.
The Korean National Police Agency set out to build a foundation that meets these requirements while letting investigation records be used safely with AI.
Platform overview
The platform is an investigation support AI system that turns investigation records scattered across documents and audio into structured data that AI can use, that is, AI-ready data.
DocuLLM, Cogniterm’s document-specialized AI, reads document images, analyzes their structure, and extracts key information. Audio records are converted to text and handled the same way as documents.
The extracted data is de-identified and pseudonymized in the Multi-modal Guardrails Engine before use, and DocuSafety manages security policies and processing history in one place. The AI Agent API Engine connects to existing work systems through REST APIs, so staff can use AI capabilities without changing how they work. All models are served on-premises, so no records leave the organization.
Key capabilities
01 Document structure analysis and information extraction AI
- Converts scanned documents and images to text with AI OCR
- Preserves document formats by analyzing layouts, tables, and key-value structures
- Automatically extracts key information such as people, dates and times, places, and amounts
- Uses the structured results to support more accurate search and analysis
02 Speech-to-text AI (STT)
- Automatically converts audio such as interviews and statements to text
- Cuts repetitive transcription work so staff can focus on review
- Searches and analyzes the converted text within the same framework as documents
03 Personal information de-identification AI
- Automatically detects personal and sensitive information in documents and transcribed audio
- Reduces exposure of personal information during AI use through de-identification and pseudonymization
- Protects the originals while keeping the data usable
04 Work system integration AI API
- Provides OCR and parsing as REST APIs that integrate with existing work systems
- Runs without any external connection thanks to on-premises model serving
- Calls AI at the right moment within existing workflows
Operations
The design centers on security and auditing for the policing environment. Given the special nature of handling investigation records, data protection and operational control were the top priorities.
- Personal and sensitive information is de-identified and pseudonymized before it reaches AI processing
- Security policies for inputs, outputs, and attachments are managed centrally in DocuSafety
- Processing and usage history is stored in audit logs for traceability
- Access is controlled through per-user permissions and stored usage logs
- AI organizes the results; staff make the final check and judgment
The platform was delivered on-premises so it runs reliably even on an internal network separated from the outside.
Expected impact
The core of this project is a data foundation that lets AI work safely with investigation records. Turning scattered document and audio records into structured data eases the burden of organizing records, so staff can focus more on reviewing facts and making judgments.
Because it combines de-identification with audit log management, this setup can serve as a reference model for other policing and law enforcement agencies that handle sensitive information.
Building on this AI-ready data, the platform can also expand step by step into investigation support AI services such as searching case records, analyzing similar cases, and drafting reports.