Proof · Trust Evidence

Proof, not promises.

Data that is ready for AI, works without sending the original out, and holds steady across runs. The cases, certifications, patents, awards and listings behind it, in one place.

88.55% Weighted F1 (Syntitan) Insurance call classification · 58.55% → 88.55% (+30pp) · 64 classes · Jan 2025 · self-measured
0.12s Per-page processing (LLM Capsule) 2,200-character page · answer generation excluded · self-tested
0.94 PII detection F1 (LLM Capsule) automated PII detection · self-tested
Customers

Across banking, insurance, the public sector, defense, and telecom.

Gartner® “Emerging Tech: AI Vendor Race: Tech Innovators in Agentic AI — Solution Accelerators” (2026)

Gartner® “Emerging Tech: AI Vendor Race: Most Prominent Use Cases in Agentic AI by Industry” (2026)

Gartner® “Emerging Tech: Provider Differentiation Strategy—Trends for Hyper-Synthetic Data” (2025)

Named a Representative Vendor · Hyper-Synthetic Data (2025)

Operational Evidence

Same model, same criteria, only the data changed.

Cases where nothing moved are recorded too. Self-measured actual records and representative examples are labeled apart.

See whether your data is in a state to start AI at all.

Representative exampleFinancial Services

Retraining pipeline: schema change caught before the next run

AI: Retraining pipeline model Data: Upstream feature tables

Held constantUse caseRetraining pipeline operationModelSame model and pipeline, version fixedEvaluationSame validation procedureEnvironmentSame runtimeReproducePre-change version re-run
Ready for Raw DataProblemWhere the data stands today
An upstream schema change caused silent degradation. Finding the root cause took a long time, and downstream decisions were affected in the meantime.
Ready for ModelWhat we applyPrepared and validated for the use case
SyntitanInput data was kept as versions, and the schema was compared with the previous version at ingestion. Two missing feature columns and one type change surfaced there.
  • Inputs kept as versions
  • Schema compared with the previous version at ingestion
Check the resultResultUsable for this use case, or not
Criteria met forthe retraining pipelineRoot causeBeforeHard to traceAfterChange found at onceChangedInput data state before and after the schema changeThe change surfaced before the next training run, so no degraded model reached production. The root cause was found in the same place.
Track and reuseOperate and recordRecorded and revalidated in operation
The pre-change version was re-run on the same pipeline and model. Every later run records which data version it used.State Card · Change Log · Re-run Record · Schema Diff

Related readingSchema Changes Break Production AI: The Silent Release

Representative exampleTelecom

Real-time inference: score drift traced and rolled back within 2 hours

AI: Real-time inference model Data: Preprocessed feature stream

Held constantUse caseReal-time inference serviceModelSame inference modelEvaluationSame production score metricEnvironmentSame serving environmentReproducePre-change state re-run
Ready for Raw DataProblemWhere the data stands today
Production scores became erratic after a preprocessing update. There was no way to trace which version the drift started from.
Ready for ModelWhat we applyPrepared and validated for the use case
SyntitanEvery run was recorded with the data version it used. Comparing it with the previous version pointed to a normalization change and a 12% shift in feature scaling range.
  • Scores recorded with the data version per run
  • Change located by comparing versions
Check the resultResultUsable for this use case, or not
Criteria met forreal-time inferenceTrace and rollbackBeforeNot possibleAfterUnder 2 hoursChangedData state before and after the preprocessing updateRolled back to the pre-drift version within 2 hours, with a record of which run used which data version.
Track and reuseOperate and recordRecorded and revalidated in operation
Each run keeps its score with the data version it used, so the next update shows at once which version changed the result.State Card · Change Log · Re-run Record

Related readingPreprocessing Drift in Production AI

Actual recordInsurance

Insurance call classification: thin training data supplemented in Syntitan

AI: Call classification model, 64 classes Data: Call records and class labels

Held constantUse caseCall classification, 64 classesModelSame classification modelEvaluationSame evaluation criteriaComparedOriginal only vs supplemented
Ready for Raw DataProblemWhere the data stands today
Some classes had too few training examples, and the model trained on the original data alone reached a weighted F1 of 58.55%.
Ready for ModelWhat we applyPrepared and validated for the use case
SyntitanSyntitan supplemented the thin classes with synthetic data, and the same model was retrained on the original data plus the supplement.
  • Thin classes supplemented in Syntitan
  • Compared against the original-only run on the same criteria
Check the resultResultUsable for this use case, or not
Same evaluation criteriaWeighted F1 improved, self-measuredWeighted F1+30ppBefore58.55%After88.55%ChangedTraining data with the original only, and with the supplement addedWeighted F1 rose to 88.55%, self-measured.
Next · Track and reuseNext stepCan be revalidated under the same conditions
The next training cycle can be re-checked in Syntitan under the same conditions.In Syntitan · Dataset Version · Re-run Record
MeasurementData6,874 call transcriptsTest set687 original recordsClasses improved55 of 64, 29 from zeroMacro F127.9% → 70.7%MeasuredJanuary 2025, self-measured

Class labels were assigned with an LLM, and these are first-round comparison results. Weighted F1 averages across classes by their data volume.

Related readingSynthetic Data Augmentation: When It Helps Classifier Training

Representative exampleTelecom

Churn prediction model: the data checked before swapping the model

AI: Churn prediction model (XGBoost v1.3) Data: Subscription and usage history, too few churn cases

Held constantUse caseChurn predictionModelSame model (XGBoost v1.3)EvaluationSame evaluation set, F1/Recall/PrecisionEnvironmentSame runtimeReproduceRe-run verified
Ready for Raw DataProblemWhere the data stands today
The churn model was missing its recall target. The team wanted to know whether the data, not only the model, was involved.
Ready for ModelWhat we applyPrepared and validated for the use case
SyntitanAdded the churn cases that were too few and restored the missing tenure context. Everything else untouched, re-run under the same conditions.
  • Missing churn cases added
  • Missing tenure context restored
Check the resultResultUsable for this use case, or not
Criteria met forchurn predictionF1+0.07Before0.75After0.82ChangedData state before and after adding churn casesRecall 0.71 to 0.84, Precision 0.79 to 0.80. Improved, and the criteria (F1 at least 0.80, Recall at least 0.82) were met. Improvement and criteria are judged separately.
Track and reuseOperate and recordRecorded and revalidated in operation
The validated data is saved as a version and every run records which version it used. When churn patterns shift, only that scope is revalidated.State Card · Dataset Version · Re-run Record

Related readingYour Model Changed and Your Data Changed. Which One Drove the Result?

Representative exampleHealthcare

Clinical AI validation: back on track without the patient records it could not access

AI: Clinical prediction model, validation stage Data: Patient records, not accessible

Held constantUse caseClinical model validationModelSame validation pipeline and modelEvaluationSame validation metricsEnvironmentSame validation environmentReproduceReproducible from the saved version
Ready for Raw DataProblemWhere the data stands today
Regulatory constraints blocked access to the patient records needed for validation, so the pipeline stalled.
Ready for ModelWhat we applyPrepared and validated for the use case
DTSThe inaccessible records were replaced with DTS reconstruction under DP. Distributional properties are preserved, with no mapping back to individuals.
  • Inaccessible records replaced with DP reconstruction
  • Distribution kept, never mapped back
Check the resultResultUsable for this use case, or not
Criteria met forclinical model validationValidation pipelineBeforeHaltedAfterRunningChangedReconstructed data in place of the patient recordsThe validation pipeline resumed without modification, with a DP audit log kept for review.
Track and reuseOperate and recordRecorded and revalidated in operation
The reconstructed data version and DP audit log stay together, so the same validation pipeline can run again at any time.State Card · DP Audit Log · Dataset Version

Representative exampleFinancial Services

Claims processing LLM: the same task, without sending the original out

AI: Claims assistance LLM, external API Data: Claims documents with names, IDs and medical details

Held constantUse caseClaims processing assistanceModelSame external LLM and promptEvaluationSame output comparisonEnvironmentMapping layer inside the customer environmentReproduceSubstitution and restoration logged per run
Ready for Raw DataProblemWhere the data stands today
Claims documents with policyholder names, IDs and medical details were going to an external LLM API as they were. Compliance halted the workflow.
Ready for ModelWhat we applyPrepared and validated for the use case
LLM CapsuleLLM Capsule substitutes identifiers such as names and IDs and keeps the context the claim decision needs. Originals and the mapping stay in a protected mapping layer inside the customer environment.
  • Identifiers substituted, claim context kept
  • Original values restored in the output
Check the resultResultUsable for this use case, or not
Criteria met forclaims processing assistanceExternal LLM pathBeforeHaltedAfterRunning on substitutesChangedA substituted working version sent instead of the originalThe external LLM path reopened. Outputs came back with the original names, codes and values restored and fed downstream as they were.
Track and reuseOperate and recordRecorded and revalidated in operation
Substitution and restoration logs are kept per run. When the substitution policy changes, only that scope is revalidated.State Card · Substitution Log · Mapping Record · Re-run Record

Related readingLLM Capsule vs AI DLP: Block the Data or Finish the Work

Representative exampleRetail

Recommendation engine: pre-upgrade state reproduced within 3 hours

AI: Recommendation model Data: User and item embeddings

Held constantUse caseRecommendation model operationModelSame recommendation modelEvaluationSame evaluation dataComparedRuntime before and after the upgradeReproducePre-upgrade state re-run
Ready for Raw DataProblemWhere the data stands today
Recommendation scores dropped after a routine infrastructure upgrade and engineers could not reproduce the pre-upgrade behavior.
Ready for ModelWhat we applyPrepared and validated for the use case
SyntitanRuntime parameters were kept as a version, and each run was recorded with it. Comparing versions pointed to a library version change and a floating-point handling difference.
  • Runtime parameters kept in the version
  • Each run recorded with the version it used
Check the resultResultUsable for this use case, or not
Reproducedthe pre-upgrade stateRoot causeBeforeDaysAfterUnder 3 hoursChangedRuntime only. The data state stayed the sameThe pre-upgrade version was re-run within 3 hours and the score gap confirmed.
Track and reuseOperate and recordRecorded and revalidated in operation
The runtime details stay in the version, so the previous state can be re-run as it was at the next upgrade.State Card · Runtime Record · Change Log · Re-run Record

Related readingHow to Reproduce an AI Incident: A 6-Step Playbook

Representative examplePublic Sector

LLM adoption under N2SF: started without sending internal business context out as it was

AI: External LLM for work under N2SF rules Data: Work documents with internal business context

Held constantUse caseLLM tasks under N2SFModelSame external LLMEvaluationPer-request output verificationEnvironmentMapping inside the boundary, local logsReproduceRe-verifiable per request
Ready for Raw DataProblemWhere the data stands today
In the segregated network the original context could not leave, so bringing an external LLM into the work had stalled.
Ready for ModelWhat we applyPrepared and validated for the use case
LLM CapsuleLLM Capsule consistently substitutes sensitive terms before processing and keeps the context the task needs. Only the substituted version reaches the external LLM. Configured with reference to the AI linkage controls in N2SF Model 2. What may go out is set by the agency's classification and security review.
  • Sensitive terms substituted, task context kept
  • Workflow built with reference to N2SF Model 2 controls
Check the resultResultUsable for this use case, or not
Example result forLLM tasks under N2SFLLM adoptionBeforeHaltedAfterIn progressChangedA substituted working version sent instead of the originalLLM adoption moved forward. Original values are restored in the output inside the company.
Track and reuseOperate and recordRecorded and revalidated in operation
Local logs are kept per request for later verification. Originals and the mapping stay inside the boundary.Local Mapping Layer · Audit Log · N2SF Reference

Related readingConfidential Business Context: The Data Beyond PII That Blocks AI

Representative exampleManufacturing

OT network data: usable for threat analysis AI with its structure intact

AI: Threat analysis AI Data: OT/ICS network data

Held constantUse caseThreat analysisModelSame analysis AIEvaluationSame threat question setEnvironmentExisting OT analysis platformReproduceFixed data state re-run
Ready for Raw DataProblemWhere the data stands today
OT/ICS network data carried sensitive operational detail and could not go to an external AI for automated threat analysis.
Ready for ModelWhat we applyPrepared and validated for the use case
LLM CapsuleSensitive fields substituted while topology and relationships stay intact (structure-preserving substitution). Runs inside the existing OT analysis platform.
  • Sensitive fields substituted, topology and relations kept
  • Runs inside the existing OT analysis platform
Check the resultResultUsable for this use case, or not
Criteria met forthreat analysisAI threat analysisBeforeBlockedAfterIn progressChangedSensitive fields substituted, relationship structure keptThe analysis AI now answers threat questions on the substituted network data, with the relationship structure preserved.
Track and reuseOperate and recordRecorded and revalidated in operation
Data states before and after substitution are fixed, and the analysis was re-run to confirm the same result.Transformed Dataset · Analysis Log · Structure Map

Representative exampleTelecom

Network operations model: topology change confirmed as the cause by re-run

AI: NOC operations model Data: Network configuration and state logs

Held constantUse caseNOC operationsModelSame NOC modelEvaluationSame output stability metricComparedConfiguration before and after the changeReproducePre-change replay, re-verified after the fix
Ready for Raw DataProblemWhere the data stands today
NOC model outputs became unstable after a network configuration change. The model itself was unchanged, so the cause was unclear.
Ready for ModelWhat we applyPrepared and validated for the use case
SyntitanNetwork state was kept as a version per run. Comparing it with the previous run identified the changed configuration elements.
  • Network state kept as a version per run
  • Configuration change identified by comparing runs
Check the resultResultUsable for this use case, or not
Criteria met forNOC operationsDrift causeBeforeUnknownAfterIdentifiedChangedState before and after the network configuration changeRe-running the pre-change state confirmed the topology change as the cause. Verified again after the fix.
Track and reuseOperate and recordRecorded and revalidated in operation
Every configuration change leaves a comparison record, so the next change can be re-run and checked the same way.State Card · Topology Diff · Change Log · Re-run Record

Related readingHow Diff Works on Data State: See What Changed in an AI Run

Representative exampleCriteria not met after refinementRetail

Pricing recommendation model: refining the data did not reach the criteria

AI: Pricing recommendation model Data: Sales and price history with gaps and outliers

Held constantUse casePricing recommendationModelSame pricing modelEvaluationSame evaluation set and hit rateEnvironmentSame runtimeReproduceRe-run verified
Ready for Raw DataProblemWhere the data stands today
The hit rate was below target in production and data quality was the working hypothesis.
Ready for ModelWhat we applyPrepared and validated for the use case
SyntitanThe same model was re-run on the same evaluation set with missing values and outliers fixed.
  • Re-run on a data state with gaps and outliers fixed
  • Model and evaluation set fixed
Check the resultResultUsable for this use case, or not
Criteria not metThreshold not met. This refinement did not reach the 65% target.Hit rate+1ppBefore58%After59%ChangedData state before and after fixing missing values and outliersHit rate 58% to 59%, short of the 65% target. The remaining failure cases and evaluation evidence guide the next refinement or AI setup review.
Track and reuseOperate and recordRecorded and revalidated in operation
The re-run record with model and evaluation set fixed stays, so the next hypothesis starts from the same conditions.State Card · Re-run Record

Related readingWhat Is Data Quality for AI? Why a Clean Score Is Not a Passing Result

Representative exampleRetail

Consolidated customer data: state checked before choosing any AI

AI: Not chosen yet Data: Customer table merged from three systems

Held constantUse caseNot set (before adoption)ModelNot chosenEvaluationSix-axis overall scoreEnvironmentSyntitan Ready for Raw DataReproduceSame file re-diagnosed before and after
Ready for Raw DataProblemWhere the data stands today
There was a plan to use customer data from three systems for AI, but no one could say in numbers what state it was in.
Ready for Raw Data · fixesRefineImprove only what came out low
Syntitan Ready for Raw DataA six-axis diagnosis on one file. Consistency (mismatched code schemes) and Traceability (no source column) came out low, so only those two were fixed.
  • Six-axis diagnosis on one file
  • Only Consistency and Traceability fixed
Data Readiness ScoreResultShared readiness score
Ready for Raw DataUse case verdict comes next in Ready for ModelOverall score · six axes+21Before58After79ChangedData state before and after refinementThis is shared readiness, not the success odds of a specific AI. Once the use case is set, Ready for Model starts from this version.
Next · Ready for ModelNext stepValidate once the use case is set
Assessment results and improvement history are kept as one version. When the use case is set, fit validation starts from it.Data Readiness Report · Dataset Version

Related readingAI Readiness Assessment for Enterprise Data: Six Axes and a Data Readiness Score

Representative exampleManufacturing

Equipment sensor logs: readiness checked before picking a predictive maintenance AI

AI: Not chosen yet, predictive maintenance candidates under review Data: 18 months of line sensor logs

Held constantUse caseNot set (predictive maintenance candidates)ModelNot chosenEvaluationSix-axis overall scoreEnvironmentSyntitan Ready for Raw DataReproduceRe-diagnosed before and after
Ready for Raw DataProblemWhere the data stands today
Predictive maintenance AI was under review, but the sensor logs had missing stretches and series in mixed units.
Ready for Raw Data · fixesRefineImprove only what came out low
Syntitan Ready for Raw DataIntegrity and Reproducibility came out low on the six-axis diagnosis. Only the gap markers and unit normalization were handled first.
  • Gap markers and unit normalization
  • Only Integrity and Reproducibility fixed
Data Readiness ScoreResultShared readiness score
Ready for Raw DataUse case verdict comes next in Ready for ModelOverall score · six axes+23Before51After74ChangedData state before and after refinementA starting point where any vendor model can be compared from the same state. The score is shared readiness, not a performance prediction.
Next · Ready for ModelNext stepValidate once the use case is set
The refined logs are fixed as one version. Once a candidate model is chosen, the same-condition comparison runs on it.Data Readiness Report · Dataset Version

Related readingAI Readiness Assessment for Enterprise Data: Six Axes and a Data Readiness Score

Representative examplePublic Sector

Citizen inquiry records: text state diagnosed before reviewing an LLM

AI: Not chosen yet, inquiry-response LLM under review Data: Inquiry text records by department

Held constantUse caseNot set (inquiry-response LLM under review)ModelNot chosenEvaluationSix-axis overall scoreEnvironmentSyntitan Ready for Raw DataReproduceRe-diagnosed before and after
Ready for Raw DataProblemWhere the data stands today
Record formats differed by department and many entries lacked category codes and outcomes. There was no basis to judge whether an LLM could be attached.
Ready for Raw Data · fixesRefineImprove only what came out low
Syntitan Ready for Raw DataContext and Usability came out low on the six-axis diagnosis. Only category codes and format normalization were handled first.
  • Category codes filled, formats normalized
  • Only Context and Usability fixed
Data Readiness ScoreResultShared readiness score
Ready for Raw DataUse case verdict comes next in Ready for ModelOverall score · six axes+25Before46After71ChangedData state before and after refinementThis score is shared readiness. Whether the data works for inquiry response is judged separately in Ready for Model.
Next · Ready for ModelNext stepValidate once the use case is set
Assessment and improvement history are versioned. When the inquiry-response use case is confirmed, fit validation uses that version.Data Readiness Report · Dataset Version

Related readingAI Readiness Assessment for Enterprise Data: Six Axes and a Data Readiness Score

Public dataset experiments

Public datasets, tested in the open.

We prepared UCI public datasets with Syntitan and compared F1 before and after on the same model settings and the same evaluation records. Each result comes from one data split; false-positive changes and limits are written up in each post.

F1 · same evaluation records BeforeAfter Syntitan
  1. Diseased tree detectionUCI Wilt · Logistic RegressionF1 0.11 → 0.77 Read
  2. Auction shill bidding detectionUCI Shill Bidding · Logistic RegressionF1 0.91 → 0.96 Read
  3. Company bankruptcy predictionUCI Polish Companies Bankruptcy · Random ForestF1 0.41 → 0.61 Read
  4. Telecom customer churn predictionUCI Iranian Churn · Logistic RegressionF1 0.56 → 0.72 Read
  5. Term deposit subscription predictionUCI Bank Marketing · Random ForestF1 0.45 → 0.61 Read
Recognition

Verified by third parties, backed by patents and research.

International standards and software quality certifications, government program selections and industry awards, and the patents and founding-team papers behind Syntitan, LLM Capsule and DTS, in one place.

By product

Standards and quality certifications

Third-party validated certifications across information security, AI management, and software quality.

ISO 27001 · Information Security Management

Information Security Management

ISO/IEC 27001:2022 · 2026

International standard for information security management. Demonstrates a systematic approach to managing sensitive information.

ISO 42001 · AI Management System

AI Management System

ISO/IEC 42001:2023 · 2026

International standard for AI management systems. Demonstrates responsible AI governance and risk management.

GS Certification Grade 1 · DTS (2025)

GS Grade 1 · DTS

GS Certification Grade 1 · 2025

Korean SW Quality Certification, Grade 1 (2025). Verified quality, eligible for public procurement.

GS Certification Grade 1 · LLM Capsule v2.0 (2026)

GS Grade 1 · LLM Capsule v2.0

GS Certification Grade 1 · 2026

Korean SW Quality Certification, Grade 1, for LLM Capsule v2.0 (2026).

GS Certification Grade 1 · LLM Capsule (2024)

GS Grade 1 · LLM Capsule

GS Certification Grade 1 · 2024

Korean SW Quality Certification, Grade 1 (2024).

KISA Fast Track 2024

KISA Fast Track · LLM Capsule

KISA · 2024 · valid through 2026-11-21

LLM Capsule received the KISA Information Security System Fast Track confirmation (2024).

AI Product & Service Verification · DTS (TTA, 2026)

AI Product & Service Verification · DTS

TTA · 2026

DTS received AI Product & Service Verification from the Telecommunications Technology Association (TTA) in 2026.

AI Product & Service Verification · LLM Capsule (TTA, 2026)

AI Product & Service Verification · LLM Capsule

TTA · 2026

LLM Capsule received AI Product & Service Verification from the Telecommunications Technology Association (TTA) in 2026.

AI Product & Service Verification · Syntitan (TTA, 2026)

AI Product & Service Verification · Syntitan

TTA · 2026

Syntitan received AI Product & Service Verification from the Telecommunications Technology Association (TTA) in 2026.

Awards and selections from government and industry

From government program selections to industry awards at home and abroad: third-party validation of our technology and business.

Deutsche Telekom T-Challenge 2026 · 2nd Place
Industry Award

Deutsche Telekom T-Challenge 2026 · 2nd Place

T-Mobile / Deutsche Telekom · 2026

Placed 2nd in the T-Challenge global open-innovation program (T-Mobile / Deutsche Telekom), recognized for LLM Capsule's context-preserving substitution and restoration of original values.

2026 Emerging AI+X Top 100
Industry Recognition

2026 Emerging AI+X Top 100

Korea AI Industry Association · 2026

Selected for the 2026 Emerging AI+X Top 100 for its AI-Ready data technology.

Global TIPS 2026
Government Program

Global TIPS 2026

Ministry of SMEs and Startups · 2026

Korea's flagship program linking private investment with government R&D for startups with proven global potential. Selected for Global TIPS 2026 (commercialization and R&D tracks) to advance Syntitan and expand overseas.

Innovative Product · DTS (Public Procurement Service, 2026)
Public Procurement

Innovative Product · DTS

Public Procurement Service · 2026

Selected as an Innovative Product by the Public Procurement Service (September 2026).

Gyeonggi Promising SME 2026 (Gyeonggi Province)
Government Program

Gyeonggi Promising SME

Gyeonggi Province · 2026

Selected as a Gyeonggi Province Promising SME (2026).

Intellyx Digital Innovator Award 2026
Industry Award

Intellyx Digital Innovator Award 2026

Intellyx · 2026
Global ICT Future Unicorn (ICT GROWTH) 2025
Government Program

Global ICT Future Unicorn (ICT GROWTH)

Ministry of Science and ICT · 2025

Selected for the Ministry of Science and ICT Global ICT Future Unicorn program (ICT GROWTH) and given a ministerial certificate plaque.

Ultra-Gap Startup 1000+ (DIPS 1000+)
Government Program

Ultra-Gap Startup 1000+ (DIPS 1000+)

Ministry of SMEs and Startups · KISED · 2025

Selected in 2025 as a top deep-tech startup (AI / big-data) in the Ultra-Gap Startup 1000+ project.

Innovative Product · LLM Capsule (Public Procurement Service, 2025)
Public Procurement

Innovative Product · LLM Capsule

Public Procurement Service · 2025

Designated an Innovative Product by the Public Procurement Service. Public institutions can adopt LLM Capsule directly through the Innovation Marketplace.

Pilot Purchase Program · LLM Capsule (2025)
Government Program

Pilot Purchase Program · LLM Capsule

Ministry of SMEs and Startups · Korea SMEs and Startups Distribution Agency · 2025

Selected for the Ministry of SMEs and Startups' Technology Development Product Pilot Purchase Program, letting public institutions purchase and validate LLM Capsule directly.

Human Technology Award 2025 · Excellence Prize, User Category
Industry Award

Human Technology Award 2025 · Excellence Prize

Hankyoreh · User Category · 2025

Excellence Prize in the User Category at the Human Technology Awards, which recognize human-centered technology.

Startup World Cup 2025 Finalist
Industry Award

Startup World Cup 2025 Finalist

Pegasus Tech Ventures · 2025

Advanced to the Startup World Cup 2025 Grand Finale in San Francisco as a finalist (October 2025).

Minister of Health and Welfare Award, AI EXPO KOREA 2025
Government Award

Minister of Health and Welfare Award · AI EXPO KOREA 2025

AI EXPO KOREA · AI healthcare innovation category · 2025
Challenge! K-Startup 2024 Excellence Award
Government Award

Challenge! K-Startup 2024 · Excellence Award

Ministry of SMEs and Startups and nine other ministries · 2024

Excellence Award in the startup league finals of Challenge! K-Startup 2024, a startup competition run jointly by ten ministries.

COMEUP 2024 · The Stars of Stars
Industry Award

COMEUP 2024 · The Stars of Stars

Ministry of SMEs and Startups · Korea Startup Forum · 2024

Named The Stars of Stars and a Global Media Pick at COMEUP Stars 2024.

Information Security Product Innovation Award · DTS (Minister of Science and ICT Prize, 2024 H2)
Government Award

Information Security Product Innovation Award · DTS

Ministry of Science and ICT · 2024 H2

Grand Prize, Information & Physical Security category, at the 2024 H2 Information Security Product Innovation Awards (Minister of Science and ICT Prize).

Information Security Product Innovation Award · LLM Capsule (Minister of Science and ICT Prize, 2024 H1)
Government Award

Information Security Product Innovation Award · LLM Capsule

Ministry of Science and ICT · 2024 H1

Winner at the 2024 H1 Information Security Product Innovation Awards (Minister of Science and ICT Prize). CUBIG won in both halves of 2024.

NextRise Global Innovator (2024)
Industry Award

NextRise Global Innovator

2024

Selected as a NextRise Global Innovator.

Patents behind CUBIG technology

Registered patents behind Syntitan, LLM Capsule and DTS. As of October 2026: 19 patents (5 registered, 14 pending) and 3 registered software copyrights. Published applications are listed with their numbers; unpublished ones are counted only.

▸ Patent · KR US Registered

AI-Based Service Providing Method Without Leaking Private Information and Client Apparatus

KR Reg. No. 10-2757651 (App. 10-2023-0133086, Registered 2025-01-16) · US Pat. No. 12,737,496 B2 (App. 18/908,054, Filed 2024-10-07, Registered 2026-09-15 · Pub. US 2025/0117517 A1)

Core LLM Capsule patent. Method and client apparatus for AI services without exposing private information. Registered in Korea (2025-01) and the US (2026-09).

View patent →
▸ Patent · KR Registered

Method and Data Processing Apparatus for De-identifying Data While Preserving Target Characteristics

KR Reg. No. 10-2926046 · App. No. 10-2023-0167085 · Registered 2026-02-06

Core DTS patent. Method for transforming source data while preserving target characteristics such as statistical distributions and label structure.

View patent →
▸ Patent · KR Registered · US Pending

Synthetic Data Generation Method Without Leaking Target Information and Client Apparatus

KR Reg. No. 10-2818137 (App. 10-2024-0017564, Registered 2025-06-04) · US App. No. 19/039,319 · Pub. No. US 2025/0252156 A1 (under examination)

DTS synthesis patent. Client-server architecture for generating synthetic data without exposing target information.

View patent →
▸ Patent · KR Registered · US Pending

Method and Data Processing Apparatus for Generating a Synthetic Dataset Containing Multiple Attributes

KR Reg. No. 10-2818136 · App. No. 10-2024-0131551 · Registered 2025-06-04 · US Pub. No. US 2026/0017275 A1 (under examination)

DTS multi-attribute synthesis patent. Method for generating complex synthetic datasets that span multiple feature columns and attribute types.

View patent →
▸ Patent · KR Pending · US Pending

Method and Apparatus for Generating a Device-Specific On-Device AI Model

KR App. No. 10-2026-0000037 · Pub. No. 10-2026-0112039 (2026-07-16) · US App. No. 19/442,680 · Pub. No. US 2026/0195592 A1 (2026-07-09)

Method for generating on-device AI models tailored to each device. Published in Korea and the US, under examination.

View patent →
▸ Patent · KR Pending · US Pending

Synthetic Dataset Validation Method and Data Processing Apparatus for Building Training Models

KR App. No. 10-2024-0174041 · Pub. No. 10-2026-0082983 (2026-06-05) · US App. No. 19/400,665 · Pub. No. US 2026/0148142 A1 (2026-05-28)

DTS validation patent. Method for checking whether a synthetic dataset fits its training purpose. Published in Korea and the US, under examination.

View patent →
▸ Patent · KR Pending

Synthetic Dataset Filtering Method and Data Processing Apparatus for Building Training Models

KR App. No. 10-2024-0174042 · Pub. No. 10-2026-0082984 (2026-06-05)

Method for filtering a synthetic dataset down to the data a training task needs. Published, under examination.

View patent →
▸ Patent · KR Pending

Method for Building a Deep Learning Model Robust to Personal Information Leakage and Inference Apparatus

KR App. No. 10-2023-0074745 · Pub. No. 10-2024-0175088 (2024-12-19) · Co-filed with Ewha Womans University

Method for building deep learning models that are hard to extract personal information from. Published, under examination.

View patent →
▸ Patent · KR Pending

Method for Building an AI Model That Processes Heterogeneous Datasets and Analysis Apparatus

KR App. No. 10-2023-0013029 · Pub. No. 10-2024-0120404 (2024-08-07) · Co-filed with Ewha Womans University

Method for building AI models that learn from datasets with different formats. Published, under examination.

View patent →

Research papers behind the products

Selected publications by CUBIG founders, from peer-reviewed venues to a survey preprint. The privacy and robustness research behind Syntitan, DTS, and LLM Capsule.

Publication · JMLR 2025

Regularizing Hard Examples Improves Adversarial Robustness

Hyungyu Lee, Saehyung Lee, Ho Bae, Sungroh Yoon · Journal of Machine Learning Research · 2025

Adversarial robustness method that regularizes hard examples to improve robust generalization.

Publication · ICLR 2024

DAFA: Distance-Aware Fair Adversarial Training

Hyungyu Lee, Saehyung Lee, Hyemi Jang, Junsung Park, Ho Bae, Sungroh Yoon · ICLR · Vienna, May 2024

Adversarial training method that enforces fairness across subgroups via distance-aware margin adjustment.

Publication · Sensors 2024

Evaluation of Malware Classification Models for Heterogeneous Data

Ho Bae · Sensors (MDPI) · 2024

Study of malware-classifier explainability on heterogeneous data. Existing explanations fall short, and high accuracy can give a misleading sense of security.

Publication · ESORICS 2024

VFLIP: A Backdoor Defense for Vertical Federated Learning via Identification and Purification

Yungi Cho, Woorim Han, Miseon Yu, Younghan Lee, Ho Bae, Yunheung Paek · ESORICS · 2024

First backdoor defense specialized for Vertical Federated Learning. It identifies and purifies backdoor-triggered embeddings at inference.

Publication · BIBM 2023

Privacy-Preserving Publishing of Individual-Level Medical Data for Cloud Services

Ho Bae, Heonseok Ha, Siwon Kim · IEEE BIBM · Istanbul, Dec 2023

Formal privacy-preserving framework for publishing patient-level medical records to cloud services, with emphasis on utility preservation under strict privacy constraints.

Publication · ESORICS 2023

FLGuard: Byzantine-Robust Federated Learning via Ensemble of Contrastive Models

Younghan Lee, Yungi Cho, Woorim Han, Ho Bae, Yunheung Paek · ESORICS · 2023

Byzantine-robust federated learning that detects malicious clients via an ensemble of contrastive models, strong under non-IID data.

Publication · RAID 2023

Exploring Clustered Federated Learning's Vulnerability against Property Inference Attack

Hyunjun Kim, Yungi Cho, Younghan Lee, Ho Bae, Yunheung Paek · RAID · 2023

Reveals property-inference privacy risks in clustered federated learning.

Publication · IEEE/ACM TCBB 2022

DNA Privacy: Analyzing Malicious DNA Sequences Using Deep Neural Networks

Ho Bae, Seonwoo Min, Hyun-Soo Choi, Sungroh Yoon · IEEE/ACM Transactions on Computational Biology and Bioinformatics · 2022

Deep-learning analysis of malicious DNA sequences for security and privacy in genomic data.

Publication · BMVC 2022

MPGAN: Membership Privacy-Preserving GAN

Heonseok Ha, Uiwon Hwang, Jaehee Jang, Ho Bae, Sungroh Yoon · BMVC · London, Nov 2022

GAN training method that prevents membership inference attacks on generated data, providing formal, provable privacy properties for synthetic outputs.

Publication · ACM AsiaCCS 2022

Membership Feature Disentanglement Network

Heonseok Ha, J Jang, Y Jeong, S Yoon · ACM Asia Conference on Computer and Communications Security · 2022

Network architecture that disentangles membership-sensitive features from model representations, reducing exposure to membership inference attacks.

Publication · IEEE Access 2021

Gradient Masking of Label Smoothing in Adversarial Robustness

Hyungyu Lee, Ho Bae, Sungroh Yoon · IEEE Access · 2021

Analysis of how label smoothing induces gradient masking, a false sense of robustness that does not transfer to true adversarial settings.

Publication · IEEE TAI 2021

Learn2Evade: Learning-based Generative Model for Evading PDF Malware Classifiers

Ho Bae, Younghan Lee, Yohan Kim, Uiwon Hwang, Sungroh Yoon, Yunheung Paek · IEEE Transactions on Artificial Intelligence · Aug 2021

Adversarial generative modeling of malware evasion: learning to produce feature-space perturbations that bypass PDF malware classifiers while preserving functionality.

Publication · IEEE Access 2020

Anomaly Detection by Learning Dynamics From a Graph

Jaekoo Lee, Ho Bae, Sungroh Yoon · IEEE Access · 2020

Graph-based anomaly detection that learns system dynamics to flag abnormal behavior.

Publication · PSB 2020

AnomiGAN: Generative Adversarial Networks for Anonymizing Private Medical Data

Ho Bae, Dahuin Jung, Hyun-Soo Choi, Sungroh Yoon · Pacific Symposium on Biocomputing · Hawaii, Jan 2020

GAN-based anonymization of private medical datasets while preserving statistical utility for downstream analysis.

Publication · PSB 2019

DNA Steganalysis Using Deep Recurrent Neural Networks

Ho Bae, Byunghan Lee, Sunyoung Kwon, Sungroh Yoon · Pacific Symposium on Biocomputing · Hawaii, Jan 2019

Deep recurrent-network method for detecting hidden messages embedded in DNA sequences (steganalysis), applied to genomic data.

Preprint · arXiv 2018

Security and Privacy Issues in Deep Learning

Ho Bae, Jaehee Jang, Dahuin Jung, Hyemi Jang, Heonseok Ha, Sungroh Yoon · arXiv:1807.11655 · 2018

Comprehensive survey of attack surfaces and defenses in deep learning systems, covering adversarial examples, model extraction, and data poisoning.

Available on

Adopt CUBIG directly from cloud marketplaces.

Access CUBIG solutions through supported cloud marketplaces.

AWS Marketplace AWS Marketplace DTS · LLM Capsule
Naver Cloud Platform Naver Cloud Platform DTS
Microsoft Marketplace Microsoft Marketplace LLM Capsule
FAQ

Frequently asked questions.

Common questions on operational evidence, public experiments, recognition, and sensitive-data handling.

What is the difference between a representative example and an actual record? +

An actual record is a result CUBIG measured itself. On this page, the insurance call classification case is one: weighted F1 rose from 58.55% with the original data alone to 88.55% after Syntitan supplemented the training data (self-measured, January 2025). A representative example explains how the mechanism works and is not real customer data. Each case carries a label that tells the two apart.

How were the public dataset experiments run? +

We prepared five UCI public datasets with Syntitan and compared F1 before and after on the same model settings and the same evaluation records. On UCI Wilt (diseased tree detection), for example, Logistic Regression F1 rose from 0.11 to 0.77. Each result comes from one data split. Each experiment post covers false-positive changes, limits, and the files to rerun the comparison.

Why publish cases where the result did not improve enough? +

Evidence is only useful for a decision when it shows when data work helps and when it does not. In the pricing recommendation case, filling missing values and outliers moved the hit rate from 58% to 59%, short of the 65% target. This page records improvement and meeting the criterion separately.

Does the six-axis overall score predict AI performance? +

No. The overall score shows the shared readiness state of the data across six axes: Usability, Integrity, Context, Consistency, Reproducibility and Traceability. It is not a success probability or a performance forecast for any AI. Whether the data fits a specific task is checked separately in Ready for Model, under fixed evaluation conditions, once the use case is set.

What is operational evidence in AI systems? +

Operational evidence in AI systems is concrete, verifiable documentation that shows how an AI system behaves in production. It shows what changed between runs, what caused the difference, and whether the same conditions can be replayed. It includes before/after outcomes, state comparisons, and re-run records.

What is reproducible AI execution? +

Reproducible AI execution means a run can be replayed and verified. The conditions include data state, schema, preprocessing logic, and runtime dependencies. In Syntitan, data is kept as versions, each run records the version it used, and a past run can be re-run on that version.

Why is AI execution unstable in production? +

AI execution becomes unstable in production when execution conditions change after deployment: data schemas, preprocessing logic, dependencies, data windows. Any of these can cause AI results to drift without any change to the model itself. When execution state is not captured and fixed, it is hard to pinpoint which change caused a production issue.

Which certifications and awards does CUBIG hold? +

CUBIG holds ISO/IEC 27001 (information security management) and ISO/IEC 42001 (AI management system) certifications, GS Grade 1 software quality certification for DTS and LLM Capsule, and AI product and service verification from the Telecommunications Technology Association (TTA) for Syntitan, LLM Capsule and DTS. Awards, selections, patents and papers are listed in the recognition section of this page with the issuing body and year.

How can I adopt CUBIG products? +

You can sign up and start on your own data, or contact us for enterprise adoption. CUBIG is also available on AWS Marketplace, Microsoft Marketplace and Naver Cloud Platform.

How does CUBIG handle sensitive data in LLM workflows? +

CUBIG's LLM Capsule substitutes sensitive values with restorable stand-ins before the data crosses to an external LLM. The original values stay in the protected mapping layer inside the enterprise boundary. The LLM operates on the substituted version, and Capsule then restores the original names, codes, and values in the output inside your environment for downstream use.

Can CUBIG be deployed on-premises or air-gapped? +

Yes. LLM Capsule can be installed on-premises or in air-gapped environments, with the setup proposed to fit your environment. Detailed architecture is available on request during evaluation.

See it on your data

Change only the data. See what improves.

Start with one AI task that is not working yet. Run it on your own data and compare before and after under the same conditions.

Try it on your data Contact us