Models & training
Pretraining, CPT, SFT, reinforcement learning, Korean data pipelines, and custom model development
Explore model technologyAbout PattyWe go beyond choosing models.
We go beyond choosing models.
We build AI systems that work.
Patty connects model training, evaluation and governance, GPU inference infrastructure, and enterprise AI products as one technical system. We build AI that understands Korean language and documents, organizational rules, and security boundaries—and remains accountable in operation.
AI should propose, people should approve, systems should execute, and every step should leave verifiable evidence.
Company thesis
Enterprise AI succeeds through operational completeness, not model rankings
Connecting a capable model is only the starting point. Real work also requires knowing who made a request, what authority and sources apply, where approval is required, and how the result will be verified.
Patty does not bolt these connections onto each product. We design models, infrastructure, agents, controls, and evidence as one system so AI can work predictably inside an organization.
Beyond model responses, toward actions an organization can own.
Korea as a design condition
Why enterprise AI requires a different design in Korea
Korean is not a translation option, and Korea’s working environment cannot be reduced to a few global-product settings. Patty treats field constraints as system requirements.
- 01
Language and interaction
We design around Hangul IME composition, initial-consonant search, mixed Korean and English instructions, and context carried through long sessions.
- 02
Documents and work formats
We connect HWP and HWPX files, tables, revision histories, and formal public-sector documents to the context in which decisions are made.
- 03
Security and deployment boundaries
Closed networks, on-premises operation, data-transfer limits, role-based authority, and approval paths are architectural requirements.
- 04
Evaluation that reflects work
Short, standardized, English-centered benchmarks do not fully describe Korean long-session performance or the quality of real work artifacts. We evaluate representative tasks and failure conditions separately.
Patty does not publish unsupported superiority scores. We show how reliably Korean instructions carry through real development and work sessions using reproducible product behavior and explicit evaluation conditions.
One company, four layers
One team connects the complete AI system
Procuring every layer separately creates gaps in accountability. Patty connects training to production through shared criteria and feedback loops.
Evaluation & governance
Workload benchmarks, red teaming, policy enforcement, human approval, and execution evidence
Explore evaluation technologyInference & infrastructure
Distributed GPU serving, KV cache, MoE, multimodal inference, and on-premises operations
Explore infrastructureProducts & digital workforces
Governed agents for coding, collaboration, work execution, and organizational operations
Explore productsHow Patty builds
Operating principles that come before the technology
- 01
Korean is a design condition
We do not add translation to a finished global product. Language, documents, and interfaces are designed together from the beginning.
- 02
Controls and evidence start on day one
Authority, approval, policy, provenance, and execution records are not security add-ons left until launch.
- 03
We test failure conditions before demos
We validate denial paths, tool failure, long sessions, recovery, and change histories.
- 04
We leave systems customers can own
We transfer evaluation criteria, architecture, operating procedures, and evidence systems.
- 05
One team owns the connective tissue
Problems between models, infrastructure, products, and governance are not deferred to another supplier.
Engagement model
Start with a bounded validation. Finish with an operable standard.
Rather than promise an immediate enterprise-wide transformation, we define representative work and acceptance criteria, then confirm the evidence to proceed at every phase.
- 01
Define the work and constraints
Select a representative workload with clear frequency, data boundaries, authority, existing systems, and failure costs.
- 02
Technical and governance diagnosis
Bring model fit, data readiness, infrastructure, policy, and accountability into one decision record.
- 03
Bounded prototype
Use production-like inputs and tools while limiting change scope and defining human approval points.
- 04
Acceptance evaluation
Measure quality, latency, cost, authority violations, recovery, and evidence completeness under agreed conditions.
- 05
Deployment and handover
Transfer the operating architecture, observability, response procedures, change management, and owner training.
Output
Durable deliverables
We produce artifacts that remain inside the organization and can be reviewed after an engagement ends.
- Decision records and risk assumptions
- Re-runnable evaluation harnesses
- Deployment, authority, and data architecture
- Operating runbooks and incident procedures
- Evidence packages for audit and security review
Research & public contribution
We return operational knowledge in forms that can be publicly examined
Trust in enterprise AI is not created by explanation alone. Through protocols, evaluation methods, technical reports, and open source, Patty publishes the assumptions and conditions behind how we judge technology.
Company record
Company information you can verify
We publish facts that can be substantiated today—not inflated measures of scale.
- Legal entity
- Patty Co., Ltd.주식회사 패티
- Technical scope
- Models · Evaluation · Infrastructure · ApplicationsOne enterprise AI stack
- Products
- Mirr Code · Mirr Enterprise · Mirr Gov · Ergazo · Crew · PilotCurrent public product portfolio
- Publication languages
- 한국어 · EnglishKorean source with a faithful English mirror
- Public leadership
- Leadership profilesVerified experience and operating principles
- Public contribution
- DARI · Technical reports · Open sourceReviewable specifications and methods
Recognition
Recognition on international stages
2025 · Top 3
Global Media Award · Top 3
Leadership
Technology and operations, owned by one team
We connect global-scale AI and data-system experience with the operating discipline of Korean enterprise and public services.
Founder & CEO
노시욱 Patrick Rho
An engineer and technology executive who turns large-scale AI, data, and commerce systems into products that organizations can operate.
CTO
신도빈 David Shin
Enterprise architecture where authority, security, audit, and scalability are the starting point of design.
SVP
김태윤 Iila Kim
A service operator who connects product, operations, and organizations so complex technology works in enterprise environments.
CSO
권기석 Leo Kwon
Technology becomes a business only when the structure fits first — strategy, business model, go-to-market, and pricing designed as one coherent whole.
Careers
We are looking for people who stay with hard problems
At Patty, model training, GPU infrastructure, agent runtimes, and organizational controls are not separate problems. One team designs and operates the whole system.
Explore careers- 01Design from first principles
- 02Turn research into working products
- 03Prove technology through operation
Adoption & partnershipsContact Patty
Bring us the problem you need to solve
We work with teams evaluating product adoption, technical validation, security review, joint research, and partnerships.