I design and ship end-to-end AI systems across computer vision, generative AI, and LLM-based agents — not research demos, production systems with real users and real acceptance criteria. That means owning the full path: translating a client's business requirement into technical architecture, preparing datasets and fine-tuning models (LoRA and beyond), and getting the result running reliably on AWS, VastAI, or RunPod.
Most of the interesting problems I've worked on sit at the boundary between "the model can do this in principle" and "the client needs this to work every time" — which usually means multi-agent orchestration with LangChain/LangGraph on the generative side, and confronting domain gap and strict false-negative bars on the detection side.
I also lead: mentoring engineers and interns, running technical and model reviews, and translating between business stakeholders and the engineering team.
Leading AI sub-teams across generative and computer-vision client projects — from technical architecture and model strategy through to production deployment. Includes Flickrz, DrawMind, TryNectar, and Bloom.
Five years building and deploying early AI/ML systems — the foundation the later computer-vision and generative-AI specialization was built on.
Four production systems, reverse-chronological by how central they are to my current focus: agent orchestration and computer vision under real constraints. Flickrz, DrawMind, and TryNectar open into a full case study.
Webtoon AI image-generation platform for a Korean distributor with 60M+ users. Designed a multi-agent pipeline — a Script Writer Agent, a Script Reviewer Agent, dedicated agents for character and per-scene image prompts, and an Image Quality Supervisor Agent — with two human-in-the-loop hard gates. LoRA fine-tuning keeps each character's identity consistent across hundreds of panels.
An AI agent system that reads and reasons deeply over complex technical engineering drawings — orchestrating LLM/VLM capabilities and purpose-built tools to extract information and answer detailed questions. A production RT-DETR detection and segmentation pipeline serves as one of those tools, locating View, Note, and Table regions with mAP above 0.95 despite limited data, compute constraints, and heavy visual overlap between classes.
Production multi-agent LLM system for biological agriculture, giving farmers data-driven treatment recommendations. LangGraph state machines route crop-lifecycle and environmental data through a RAG pipeline built and validated with agronomists.
Live, profitable multimodal AI companion product — text, image, and video generation orchestrated through ComfyUI and LangGraph. Solved latency and character-consistency problems at scale via dynamic persona injection and context-window management.
Years of hands-on experience per area, out of 7 — the length of my AI/ML career so far.
Open to new roles in computer vision and applied AI. The fastest way to reach me is LinkedIn.