Jungmoo Cheon

Built agentic AI audit SaaS products that cut a multi-hour-to-multi-day manual audit to ~3 minutes, with 95% result agreement against expert audits and now used by domain experts directly. Meta Paid Ads audit SaaS: Architected and built an AI audit SaaS for Meta paid-ads accounts. - Orchestrated a parallel multi-agent pipeline of 8 specialized multimodal LLM agents, deploying a robust backend architecture where each agent is validated by a dedicated evaluation harness. - Decoupled deterministic analytics from the LLM reasoning layer, implementing a structured data handling system to prevent numeric hallucination and ensure enterprise-grade data integrity. - Built a RAG-powered chatbot for conversational Q&A over audit results. Organic Social audit SaaS: Designed and shipped the end-to-end LLM analysis pipeline for an organic-social audit SaaS (Instagram, YouTube, TikTok) as the sole engineer, from development to commercialization. - Collaborated directly with domain experts to extract their audit methodology, encoding it into a multi- layered, chain-of-thought prompt system via an internal prompt-management tool. - Engineered the context layer to fit large multimodal inputs (video, images, text) within model limits, splitting oversized content across multiple LLM calls and merging results without losing scoring signal. - Chained multiple LLM calls into a multi-stage pipeline with reliability layers (structured outputs, retries, fallbacks) for robust, recoverable runs. Developed secure MCP servers collapsing 25+ tool schemas into a single code-execution endpoint (model-written TypeScript in isolated microVMs) to cut token consumption. Built a Gemini fine-tuning pipeline on Vertex AI (dataset validation, tuning, evaluation) to specialize models for audit scoring.

Serialized unstructured data (PDF, LaTeX, Image) into hierarchical formats using multimodal models, increasing consistency across the pipeline and enabling efficient model ranking and evaluation through test suites. Refactored and optimized LLM-based data processing pipelines to enhance scalability and maintainability. Applied prompt engineering for educational content extraction and generation, integrating evaluation frameworks with Promptfoo and a custom semantic comparison algorithm. Built internal Streamlit tools and a data integrity system to manage test suites, ensure reproducibility, and streamline model evaluation & comparison. Researched, benchmarked, and fine-tuned Hugging Face models for object detection and OCR, achieving measurable improvements in model accuracy.

- Creation of Deep Learning Bootcamp Designed a 5-module bootcamp enabling practical, hands-on experience in building discriminative deep learning models from scratch in Python, covering essential deep learning fundamentals. - Research & Case Studies on Modern AI Architectures Conduct in-depth research and case studies on state-of-the-art AI models.

Research in LED semiconductors, Collaborated with senior researchers to enhance performance, reduce defect rates, and test new LED structures.