AI Engineer / NLP Engineer

LLM Serving Infrastructure

• • PagedAttention-based vLLM — from initial evaluation through K8s production deployment
• • V0→V1 engine migration — 84K-token EPC contracts 478s→100s (4.8x speedup)
• • RadixAttention prefix caching for repetitive contract review patterns
• • LiteLLM multi-model routing for GPU memory efficiency

Long-Context Legal Document Processing

• • 54K-80K token legal contract production pipeline
• • 'Lost in the Middle' position bias — empirical validation in legal domain via Needle-in-a-Haystack
• • Long-context SFT — base Gemma-3 hit rate 0.081→0.383 (approaching GPT-4.1's 0.390)

GPU Infrastructure & Model Training

• • AICA program — A100/H100 GPU resources, outstanding performance recognition
• • 5 task-specific LoRA adapters on Qwen2.5-14B (checklist·edit·translate·extract·Contract-AI)
• • LlamaFactory bug patch — training time 12h→3h (75% reduction)
• • 4 PRs to LoRAX OSS — serving infra behind LoRA Land paper

Structured Output & Guided Decoding

• • JSON schema, regex, FSM guided decoding — contract classification·checklist·drafting
• • response_format policy design across vLLM/LoRAX
• • LoRAX PR #644 — structured output serving interface implementation

Evaluation Metrics System

• • 16+ metrics designed — traditional 6 + G-Eval 5 + contract-split 5
• • Legal expert rubrics codified (policy compliance·revision quality·coherence)
• • GPT-4.1 / Claude Sonnet 4 / custom judge model simultaneous support
• • 10+ model benchmark leaderboard

Contract AI & Review Systems

Core product API covering contract extraction, review, and revision. FastAPI-based SSE streaming, vLLM serving, structured output. 2.5 years in production.

• + Checklist v2 human eval 80.41% (GPT-4o 79.16%)
• + 38 contract types × 5 languages
• + 4 enterprise PoC deployments
• + Korea registered patent 10-2867168
• + 54K-80K token EPC contract long-context processing

LLM Training & Alignment Pipeline

End-to-end from legal SFT dataset design to SFT/DPO training on A100/H100 GPUs and service deployment.

• + Multi-task SFT overall 0.9138 (GPT-4.1 SOTA 0.8731, +4.7%)
• + Training time 12h→3h (75% reduction, LlamaFactory bug patch)
• + 5 task-specific LoRA adapters on Qwen2.5-14B
• + Long-context SFT: hit rate 0.081→0.383 (approaching GPT-4.1's 0.390)

Workflow Platform & Agent Tooling

LangGraph-based workflow engine and agent platform.

• + 10+ workflow types in production
• + Agent Legal launch — paid workspaces 41→227 (team metric)
• + Coding Agent CLI for agentic testing

AI Evaluation & Metrics System

Legal domain-specific evaluation framework combining traditional quantitative metrics + LLM-as-a-Judge + legal expert rubrics.

• + 16+ legal AI task metrics designed and implemented (traditional 6 + G-Eval 5 + contract-split 5)
• + GPT-4.1 / Claude Sonnet 4 / custom judge model simultaneous support
• + 10+ model simultaneous benchmark leaderboard
• + Model ensemble + human review gold standard dataset

OSS: predibase/lorax

4 PRs merged into LoRAX — a production implementation of S-LoRA (MLSys 2024) and Punica (MLSys 2024) multi-LoRA serving research.

• + Contributed 2024-03 ~ 2024-10, Rust/Python
• + Directly contributed to LoRA Land paper's (310 models evaluated) serving infrastructure
• + Internal fork → K8s production serving
• + Structured output (PR #644) → internal contract classification/checklist pipeline