Youngbin J.

As an AI engineer and researcher with 8 years of experience, I design, develop, and deploy production-grade AI agents, RAG systems, and proprietary AI models for organizations ranging from idea-stage startups to Fortune 500 subsidiaries. For a startup, I developed an AI service that scaled to 1M+ users and supported a $6M+ funding raise. For an established enterprise, I developed AI agents that are running live in a production app at a $500M+ annual-revenue company. I build stable, extensible AI products that operate in real-world environments, not just demos. Whether you're starting from an idea or integrating AI into an existing service, I can help you build an AI product that works reliably in the real world and delivers value customers will pay for. 🌐 Website link youngbin-jang.pages.dev ✅ What I Actually Do I engineer the complete path from concept to production — the agentic workflows that ground your AI, and the evaluation loops that make its behavior measurable. 👉 AI Agents & LLM Workflows — Agents with memory, tool use, guardrails, and multi-agent orchestration, plus deterministic pipelines for multi-step business logic and structured output. 👉 Advanced RAG & Knowledge Graphs — Citation-backed retrieval: semantic chunking, hybrid search, and full GraphRAG — ontology design, LLM extraction with quality control, Neo4j, and text-to-Cypher agents. 👉 Custom Models & Voice AI — Fine-tuning LLMs and training proprietary models for performance, cost, and privacy constraints; domain-adapted STT, diarization, and streaming transcription on a GPU cluster. 👉 Applied AI Research — Published, state-of-the-art methods turned into production-grade services. 👉 Evaluation & Reliability — Prompt design plus eval harnesses, LLM-as-judge scoring, and hallucination detection, so quality is a number you can track. 👉 MCP Servers & Agent Harness — Your data and APIs as agent tools over MCP, plus the harness — scheduling, retries, permissions, human-in-the-loop — so agents run unattended. 👉 End-to-End Delivery — The full stack around the AI (frontend, backend, DB, deployment) — a user-ready product, not components. 🎯 Ideal Projects & Fit ✅ Strong fit: - You're building AI that people will rely on in real workflows — not just demo. - You treat launch as day one and want to refine on real usage. - You're investing in long-term stability and maintainability. 🚫 Not a fit: - A throwaway demo with no intent to operate the system afterward. 🧩 Selected Projects 👉 Proprietary Essay Scoring & Correction Models — Testglider / Databank (1M+ users, $6M+ raise) - Built, trained, and deployed in-house AI models for TOEFL/IELTS writing and speaking, replacing the LLM-based pipeline - Runned the full lifecycle: data pipeline, training, evaluation, GPU serving 👉 State-of-the-Art AI Research — Databank (arXiv:2309.02740) - Rubric-Specific Approach to Automated Essay Scoring with Augmentation Training — a state-of-the-art result - Planned, designed, executed, and authored the study — from research question to published paper 👉 GraphRAG Consulting Chatbot & Report Copilots — Jinong / Farmhannong, live in production - Domain ontology (35+ node types) and an LLM extraction pipeline with conflict resolution and an LLM-judge eval loop - Neo4j multi-database graph agent with text-to-Cypher, live at a Fortune 500 subsidiary 👉 Multi-Agent Marketing Campaign Platform — STIA, active engagement - Research → strategy → content → calendar pipeline with human review gates, concurrent execution, and live progress streaming 👉 Domain STT & Speech Assessment Stack — Testglider, Jinong (1M+ users) - Fine-tuned domain-specific ASR and diarization on a B200 GPU cluster; batch, streaming, diarized APIs - Self-hosted ASR ported to ONNX Runtime, replacing a per-call vendor STT bill 👉 AI Course-Material Generation SaaS — MangoFactory (commercial, paying users) - Built solo end to end — React/Supabase + FastAPI, agent pipeline, deployment — RAG over a 120GB+ corpus 👉 Multi-Agent Research & Report Automation — FederationAI - 10 coordinated agent crews producing export-buyer research reports, validated against real trade data 👉 LLM Fact-Checking & Hallucination-Detection SDK — Upwork, 5★ - Claim extraction, evidence retrieval, and verdict scoring, as a reusable package 🧰 Tech Stack - Agents & LLMs: Python, LangGraph, CrewAI, LangChain, MCP, Claude / GPT / Gemini - Models & Voice: PyTorch, HuggingFace, ONNX Runtime, LoRA / fine-tuning, W&B, ASR, pyannote diarization, streaming (SSE / WS) - Data: Neo4j, pgvector, Qdrant, PostgreSQL, hybrid search + reranking - Infra: FastAPI, Docker, AWS, GCP, Celery / RabbitMQ, GPU serving - Frontend: React, Next.js, Svelte, Supabase 🚀 The Next Step If you want to build AI that actually works but aren't sure where to start, click the message button and tell me what's in your head. I'll come back with an architecture sketch and an honest read on scope.