Core Strengths

• Production LLM apps for legal and business document domains • Contract review AI system design and implementation • Multi-scenario operations via workflow engine • Widget protocol and Rich UI response design

• Prompt design for legal and business domains • Optimization per task type: generation, classification, extraction • Systematic prompt management via playbook structure • Prompt templates for 38 contract types

• Playbook self-serve architecture design • Demo app planning and customer PoC prototyping • Feature planning with non-developer users in mind • Enterprise PoC execution (Samsung Biologics, etc.)

• Message-based state management and tool call orchestration • Confirmation/query pattern design (AskUserQuestion, etc.) • Workflow subgraph composition • Agent observability (Opik/Sentry)

• LangGraph-based cyclic agent graph design • 10+ legal automation scenarios in production • Tool calling and subgraph composition • SSE streaming-based real-time execution

• Reusable CPT recipe across decoder LLMs (Llama, Solar) • ~100GB legal-domain corpus curated in-house • LoRA + pretraining to control shift and overfit • General capability preserved, domain metrics rise

• 5 LoRA adapters trained on Qwen2.5-14B • Checklist, editing, translation, extraction, Contract-AI • Multi-task SFT overall 0.9138 achieved • SFT→DPO 2-stage alignment pipeline

• Transformer-based NLP model development • Document understanding, summarization, information extraction tasks • NAVER CLOVA Note title generation model • HuggingFace ecosystem hands-on experience

• LLM.int8() based OPT-13B RTX 3090 fine-tuning • INT8+LoRA — equivalent performance to FP16, 3-17% memory savings • GPTQ, mixed-precision quantization experience • Early exploration before QLoRA (2022-2023)

• JSON schema, regex, FSM-based guided decoding • vLLM structured output policy establishment • XGrammar-based 100x acceleration pipeline • LoRAX structured output open-source contribution

• Contract review, clause analysis, risk assessment systems • Registered patent (10-2867168) • 220+ PRs in production over 2.5 years • 38 contract types, 3 languages (Korean/English/Japanese)

• Document structuring, summarization, title generation, information extraction • NAVER CLOVA Note title generation model development • BHSN legal document intelligence system • Processing unstructured documents (contracts, opinions, etc.)

• 54K-80K token EPC contract serving optimization • Prefix caching for recurring pattern acceleration • Citation F1 evaluation pipeline • Needle-in-a-Haystack position bias analysis and SFT improvement

• Legal domain-specific RAG pipeline • Hybrid retrieval + reranking implementation • Citation verification and clause linking system • Query transformation for improved search accuracy

• LLM automated scoring evaluation system design • G-Eval chain-of-thought scoring implementation • Custom judge model training (SFT+DPO) • 70-80% cost reduction compared to Claude

• Designed 20+ domain-specific evaluation metrics • Korean-tokenized ROUGE, citation F1 • Structured accuracy and legal NLP metric framework • 10+ model benchmark comparison framework

• End-to-end pipeline: preprocessing → filtering → synthesis → annotation • Positive/negative filtering — 48% bad data reduction • LLM-based synthetic data and weak labeling • SFT/DPO training set construction and quality management

• Designed and built annotation UI for non-developers • Model performance and data quality visualization via evaluation dashboards • Rapid prototyping with Streamlit • Built weak label consistency verification tools

• Prefix caching, structured output, LoRA serving • Serving and testing across various GPU specs • V0→V1 migration — 84K token 4.8x acceleration • PagedAttention-based GPU memory optimization

• Multi-LoRA adapter concurrent serving via Predibase LoRAX • Shared base model + per-request adapter routing • Open-source contributions including structured output (4 PRs merged) • Migration to vLLM V1 after native LoRA support

• Multi-LLM routing platform with model layer separation • YAML config-based multi-vendor model management • Automated deployment per dev/stage/prod environments • Fallback and load balancing policy design

• ML model serving containerization • Docker Compose dev environment setup • GPU passthrough and CUDA runtime configuration • Internal service and dev environment Docker deployment

• In-house GPU server (H100x8) direct setup • System design and development for air-gapped environments • Air-gapped environment operations during military service • Client closed-network deployment experience

• GCP primary — Vertex AI, Cloud Run • AWS, Azure limited exposure • Comparative evaluation of Korean clouds (Kakao, NHN, Naver Cloud) • Docker-based internal service and dev environment deployment

• Production AI API server development (model serving, business logic) • SSE streaming-based real-time responses • X-Locale header multilingual support design • OpenAPI schema automation and documentation

• RTX 3060/3080/3090 training and inference • A5000 24GB, A6000 48GB benchmarking • INT8+LoRA single GPU fine-tuning • GPU memory profiling and optimization • NPU collaboration (Hyperaccel/Rebellions) and vendor evaluation

• A100 80GB — vLLM V1 engine 4.8x acceleration • H100 8-pack multimodal legal LLM training • AICA national compute program outstanding achievement • Large-scale model training and serving infrastructure operations

• T4 — cost-efficient INT8 inference serving • L40s — high-throughput production serving • Compute-bound vs memory-bound workload analysis • GPU selection based on model size, batch, and sequence length

• Primary language — AI/ML, backend, data pipelines • FastAPI, PyTorch, HuggingFace ecosystem experience • Async programming (asyncio, aiohttp) • Type hint-based code quality management

• Code review and PR-based collaboration workflows • Branch strategy design and CI/CD integration • GitHub Actions pipeline configuration • Open-source contributions (4 PRs merged to predibase/lorax)

• Air-gapped network security monitoring and UTM management • Security automation script development • Cyber Operations Commander commendation and InfoSec Torch Award • Zero-dependency environment operations experience

• External conference talks: LangCon 2024, Build with AI • 2 allibee engineering blog posts • 15 Confluence API specifications authored • Google ML Bootcamp Alumni presentation

• Hancom project solo lead (full Confluence documentation) • Decision logs, MVP timeline, risk management • AICA performance reports and deliverable management • Samsung Biologics PoC planning and execution

• Collaboration with product, operations, QA, and external partner stakeholders • AI feature product integration and deployment process design • Partner-specific strategies for Hancom/BoostDraft • B2G partner integration technical coordination