Models & Training
Design models and data
Technology
We publish how AI systems are actually engineered.
From model training through evaluation, inference infrastructure, and applied systems, 21 technical monographs document the decisions, methods, validation, and failure boundaries Patty uses in real engagements.
- Notes
- 21
- Domains
- 4
- Languages
- KR / EN
Four technical layers connected as one operating system
A capable model is not yet an operable AI system. Training lineage must reach evaluation; evaluated models must run on predictable infrastructure; and inference must enter work through explicit authority and evidence.
Evaluation & Benchmarks
Validate against real work
Inference & Infrastructure
Serve under operating constraints
Applied AI
Connect controlled work
Models & Training
Pretraining & CPT
A CPT discipline spanning corpus diagnosis, mixing, tokenization, distributed runs, checkpoints, and serving acceptance.
SFT Data Pipelines
Design source ingestion, schemas, deduplication, contamination control, human audit, calibration, and immutable exports.
RL Alignment: GRPO & RLHF
Design task suitability, rollouts, rewards, GRPO/RLHF choice, reward hacking controls, evaluation, and recovery together.
Training Dataset Generation
Design source taxonomy, synthetic generation, independent audit, sensitive-data removal, coverage, and human sampling.
Korean-Specialized Modeling
Address morphology, endings, spacing, token efficiency, Korean-English mixing, long sessions, and evaluation gaps together.
Custom Model Development
Design the build-versus-adapt decision, organization-specific layers, versioning, safety gates, and deployment boundaries.
Evaluation & Benchmarks
Benchmark Design & Engineering
Combine representative workloads, scoring rules, contamination controls, human review, and versioning in one evaluation protocol.
Model Evaluation & Red Teaming
Manage capability, safety, security, bias, and tool-use failures through threat models and evidence-based release gates.
Inference & Infrastructure
Distributed GPU Serving & Scheduling
Model request shapes, parallelism, batching, routing, and failure domains together to engineer service levels.
KV Cache Architecture
Design block lifecycle, prefix reuse, tiered storage, eviction, and tenant boundaries around the workload.
MoE Serving (Wide EP)
Co-design batching, routing, all-to-all communication, and failure scope for wide expert parallel deployments.
Multimodal Inference
Engineer resolution, ordering, caching, batching, and output evidence for image, document, and audio inputs.
RL Rollout Acceleration
Build a measurable lifecycle across inference workers, sample buffers, reward computation, and trainers.
GPU Farm Operations
Unify capacity planning, telemetry, workload isolation, maintenance, and recovery drills in one operating model.
Applied AI
Korean-Specialized RAG
Design the path from Korean document ingestion through structure-preserving chunking, hybrid retrieval, reranking, evidence, and access control.
Document Intelligence Pipelines
Preserve layouts, tables, forms, and signatures across HWP, HWPX, PDF, and images while linking extraction, validation, and approval.
LLM-Native Browser
Engineer DOM, accessibility, and visual state with session isolation, tool authority, approvals, and reproducible execution records.
Agent Orchestration
Design task decomposition, tool contracts, state, approvals, retries, and compensating transactions as an observable graph.
AI Workforce Transformation
Connect work decomposition, risk tiers, roles, approvals, quality measurement, and change management in an operating model.
ViewSpec Rendering
Place typed component vocabulary, data binding, accessibility rules, policy validation, and visual regression at the rendering boundary.
On-Premises Deployment
Design network boundaries, supply chain, secrets, model import, observability, patching, backup, and handover from a threat model.
Begin with the bottleneck in your current system.
We map representative workloads, system constraints, and acceptance evidence before proposing an intervention.
Request a technical review