AI decision systems
From unstructured data to operational decisions.
FairLabs builds AI decision systems for financial markets. Their foundation is the extraction and analysis of unstructured data, with every output traceable to its source.
A Korea Exchange (KRX) company
Five layers, one lineage
Documents enter at the base. Each layer adds structure and meaning until decisions are made at the top, and every output remains traceable to its source.
- Filings processed
- 191,661
- Periodic filings parsed without loss, 2010–2026
- Tables structured
- 72.4M (72,452,577)
- Kept with units, footnotes and section context
- News articles monitored
- 1.15M (1,152,604)
- Collected continuously since July 2026
- Eligibility decisions
- 840,753
- Security-level screening decisions computed
As of · Aggregate counts from FairLabs production systems
The challenge
The data that matters most is the hardest to use.
Most of the information behind a financial decision sits in documents, transcripts, tables, scans and news. Little of it arrives in structured form, and once it is copied into a spreadsheet, its provenance is lost.
- Heterogeneous formats
- Text, tables, images and audio, often within a single file.
- No common schema
- The same fact is reported in many different forms.
- No provenance
- Figures become separated from the passages that support them.
Pipeline
From document to decision, traceable at every step.
Agents parse, extract and validate in parallel. Every fact carries its source passage, its agent, its confidence and its review status, so every decision can be traced back to the page it came from.
01 · Sources
Filed document
PDF · 212 PAGES
Call transcript
AUDIO → TEXT
News feed
STREAM · 40 SOURCES
02 · Parse
Parse · OCR
LAYOUT + TABLES
03 · Extract
Extract · Events
AGENT
Extract · Entities
AGENT
Extract · Figures
AGENT
Extract · Relations
AGENT
Extract · Claims
AGENT
04 · Validate
Validate
RULES + REVIEWER
05 · Ontology
Ontology write
OBJECTS + LINKS
06 · Decision
Decision
REVIEWED · V3
- Documents
- 1,284
- Pages parsed
- 96,410
- Facts extracted
- 312,077
- Human-reviewed
- 4.1%
Illustrative pipeline and figures.
Ontology
One shared model, at every level of detail.
Extracted facts are written into an ontology of entities, events, relations and signals. Analysts and AI agents query the same model, from object types down to the evidence behind each one.
- Versioned schemaClasses and relations change only through review, and every change is recorded.
- Shared by people and agentsThe model an analyst reads is the model an agent reasons over.
- Evidence at every objectEvery object resolves to the passages that support it.
- Organization24.0%
- Announcement17.0%
- Person12.0%
- Ownership link11.0%
- Transaction9.0%
- Role link8.0%
- Anomaly signal7.0%
- Instrument5.0%
- Trend signal4.0%
- Schedule event3.0%
Example distribution.
Decisions
Every decision carries its evidence.
Each decision is recorded with the passage it read, the facts it extracted, and the rule version and review state that applied.
Index eligibility screening
d-01 · Eligibility rules v4.001Source
Annual report · periodic filing
§ VII Major shareholders · Table 2
Largest shareholder and related parties · Common shares · 71.3%
02Extracted facts
- Entity
- Company A · listed, KOSPI
- Largest-shareholder stake
- 71.3% of common shares
- Free-float ratio
- 24.6% after treasury and locked-up shares
03Decision
Excluded from the eligible universe
Free float below the index rule's minimum. Stake, exclusions and rule clause are attached to the decision.
ApprovedConfidence 0.99
Financial table structuring
d-02 · Table schema v2.401Source
Quarterly report · consolidated income statement
Income statement · Note 3
(Unit: KRW thousand) Operating profit … * Prior period restated (see Note 3)
02Extracted facts
- Unit
- KRW thousand, normalized to KRW
- Restatement
- Prior-period operating profit restated
- Footnote
- Note 3 linked to the restated row
03Decision
Prior-period series updated
The earlier value is kept as a prior version, so past analyses can still be reproduced as they ran.
ApprovedConfidence 0.98
News event monitoring
d-03 · Event model v2.101Source
News article · wire service
Headline and lead paragraph
… plans to raise new capital through a rights offering, the company said …
02Extracted facts
- Event
- Capital increase · rights offering
- Entity match
- Two listed candidates with similar names
- Linked filing
- None found yet
03Decision
Sent to a reviewer before it enters the ontology
Entity confidence is below the write threshold, so a person confirms the company before any signal acts on it.
In reviewConfidence 0.62
Illustrative examples modeled on production workflows. Company names, figures and passages are anonymized.
Operations
Designed for continuous operation.
Decision systems earn trust in production. FairLabs systems run continuously, report their own state and route uncertain cases to human reviewers.
- Running09:04 KST
Disclosure intake
r-24117 · Extract · v3.2
- In review08:57 KST
Call transcript analysis
r-24116 · Validate · v1.8
- Approved08:51 KST
Entity resolution
r-24115 · Ontology write · v2.4
- Approved08:40 KST
Anomaly signals
r-24114 · Decision · v4.0
- Running08:38 KST
News event stream
r-24113 · Extract · v2.1
| Run | Pipeline | Stage | Status | Version | Updated |
|---|---|---|---|---|---|
| r-24117 | Disclosure intake | Extract | Running | v3.2 | 09:04 |
| r-24116 | Call transcript analysis | Validate | In review | v1.8 | 08:57 |
| r-24115 | Entity resolution | Ontology write | Approved | v2.4 | 08:51 |
| r-24114 | Anomaly signals | Decision | Approved | v4.0 | 08:40 |
| r-24113 | News event stream | Extract | Running | v2.1 | 08:38 |
- 01Monitoring
- Throughput, latency and extraction quality are tracked per pipeline, with alerts when they drift.
- 02Human review
- Low-confidence facts and every decision above a set threshold go to a reviewer before they act.
- 03Versioning
- Models, prompts, schemas and rules are versioned, so any result can be reproduced as it was.
- 04Audit trail
- Who changed what, which version decided, and which passages it relied on, kept for every run.
Core capabilities
Data analysis at the core, held to production standards.
- 01Unstructured data extractionDocuments, transcripts, tables and images become structured facts, each tied to the exact passage it came from.
- 02Data analysis and modelingQuantitative and statistical analysis of the extracted data, the foundation on which the rest of the system depends.
- 03Ontology engineeringEntities, events and relations modeled so that analysts and AI agents reason from the same view of the world.
- 04Decision systems in operationAI decision systems deployed into live workflows, with monitoring, human review, versioning and a full audit trail.
Company
Built inside financial market infrastructure.
FairLabs is part of Korea Exchange (KRX). Its AI decision systems are built to the standards that environment requires: every decision auditable, every change versioned, every source retained.
A Korea Exchange (KRX) company
Documentation · available to clients
- MethodologyHow sources are ingested, how facts are extracted, and how decisions are formed.On request
- Model evaluationTest sets, metrics and results for every extraction and decision model in production.On request
- Data securityAccess control, isolation, retention and incident handling.On request
- Change logEvery model, schema and rule change, with its review and release date.On request
FAQ
Frequently asked questions
- What is FairLabs?
- FairLabs is a Korea Exchange (KRX) company based in Seoul. It builds AI decision systems for financial markets that run in real operations. Its core competency is data analysis: turning unstructured data into decisions that are reviewed, versioned and traceable to their source.
- What does FairLabs build?
- FairLabs builds AI decision systems in five layers: a data foundry that extracts facts from unstructured sources, a versioned data layer, a semantic layer of shared definitions, an ontology of entities, events and relations, and a decision layer where AI agents and analysts act in production.
- How does FairLabs work with unstructured data?
- Documents, transcripts, tables and images are parsed and chunked, and candidate facts are extracted together with the passage they came from. As of 17 October 2026, FairLabs systems have parsed 191,661 corporate filings without loss and structured 72.4 million tables with their units, footnotes and section context.
- How are FairLabs decisions kept auditable?
- Every pipeline is monitored, low-confidence results go to human reviewers, models, schemas and rules are versioned, and an audit trail records which version made each decision and which source passages it relied on.
- How is FairLabs related to Korea Exchange?
- FairLabs is a subsidiary of Korea Exchange (KRX). It builds and runs its AI decision systems to the audit, accountability and security standards of financial market infrastructure.
- How can I contact FairLabs?
- Email team@fairlabs.io. FairLabs is located at Parc.1 Tower, 108 Yeoui-daero, Yeongdeungpo-gu, Seoul.
Contact
Discuss your requirements with our team.
Describe the decision you need to support and the documents behind it. Our team will follow up with an assessment of scope and approach.
team@fairlabs.io