Skip to content

$ whoami

Seongjin Lee

LLM service engineer for specialized business domains

I've been digging into NLP and LLMs since before ChatGPT. An engineer working to turn that technical grounding into solutions for real business problems.

  • End-to-end design and development capability across model training, evaluation, serving, workflows, and agents
  • Productization that covers both backend APIs and the operational tools non-engineers use directly
  • Deep understanding of Transformer decoder LLM architectures and the tech trends around them
  • Design and development capability for understanding the contract-law domain and shaping it into a real, running product
View Projects → Get in Touch
AI Competency Stack
View Details →

Career

View all →
BHSN AI Engineer
  • Designed contract review AI system (patented)
  • Built LLM training, serving & evaluation pipelines
  • Workflow/agent platform & external partner integrations
NAVER Cloud (CLOVA) Research Intern
  • CLOVA Note title generation model development
  • SFT with automated data curation improving 4 preference types
Republic of Korea Army InfoSec Specialist
  • Security automation in air-gapped environments
  • Commendation from Cyber Operations Command

Featured Projects

View all projects →
AI Workflow & Agent Platform
  • LangGraph-based workflow engine powering 10+ legal automation scenarios
  • Supports contract drafting, comparison, review, compliance, and document inspection via tool calling and subgraph composition
→
AI Contract Review System
  • End-to-end AI system for contract analysis covering extraction, checklist review, risk assessment, and revision suggestions
  • Patent registered
→
LoRAX Open-Source Contributions
  • 4 PRs merged to predibase/lorax
  • Improved OpenAI-compatible serving interface, deterministic generation, streaming usage, and structured output support
→
Legal Domain LLM Training & Alignment
  • End-to-end training pipeline for legal domain LLMs
  • Achieved SOTA on contract analysis benchmarks using SFT, DPO, and task-specific LoRA adapters
→
Long-Context Legal Document LLM — Training, Evaluation & Serving
  • End-to-end long-context work on legal documents of 50K–100K+ tokens: successful open-model SFT convergence at 100K+ context on a single GPU server node, a matching long-document evaluation pipeline, and production serving on top of the trained model
→