Most companies still move documents by hand. Translating a 90-page manual. Pulling numbers out of a stack of PDFs. Checking whether the totals actually add up.
The tools that exist do half the job. ChatGPT translates well and wrecks your layout. DeepL and Google break multi-column pages and complex tables. Document AI from the big clouds extracts fields but won't tell you when the line items don't reconcile with the summary on page 3.
I build the other half.
What I've built and run:
- A format-preserving docx translation pipeline with translation memory, terminology control, and do-not-translate protection
- A causal-chain research pipeline for trade and tariff documents (2asy.ai), around 2,400 articles a day
- A multilingual entity-resolution API
- Live demos: PDF Q&A and PDF-to-JSON extraction at demo.hannune.ai
Before Hannune: LG Energy Solution, Scenera (Netflix, Sony, GS Retail). MS, Seoul National University. Two patents.
Automated AI-powered economic news platform with dual-language support
- 🌍 Live at 2asy.ai (English) & 2asy.net (Korean)
- 📰 10 daily automated publications powered by local LLMs
- 💰 Reduced costs to 1/10-1/20 compared to cloud APIs
- 🤖 Custom summarizer with structured output & content review nodes
- 🔄 Fully containerized pipeline with vLLM, llama.cpp & LiteLLM
Production-grade infrastructure for running 70+ local LLM models
Architecture
Your App ↓ LiteLLM Gateway ← Load balancing, fallbacks, metrics ↓ ├─────────┬─────────┐ │ │ │ Ollama Ollama vLLM ← 70+ models across servers
Components:
- 🚀 Fleet Manager - Manage 70+ models (Qwen, Llama, Granite, Mistral)
- 🌐 LiteLLM Gateway - OpenAI-compatible proxy with load balancing
- 🐳 Production Docker - Hardened Ollama deployment
- 📊 Monitoring Dashboard - Real-time GPU & model status
- ⚡ GPTQ-Marlin Optimization - 2-3x faster inference
Reusable building blocks for local LLM applications
Featured Components:
- 🔍 Elasticsearch RAG Manager - Full-stack RAG with FastAPI + Streamlit
- 🎨 Vision-Language Models - Local VLM experiments (Granite, LLaVA)
- 🔧 MCP Agent Services - Code execution, research & web scraping microservices
- 📝 Large Text Summarizer - Map-reduce for documents exceeding context windows
- 📰 GDELT Collector - Real-time global news with advanced filtering
- 🔗 Distributed Tools - MCP + LangChain integration
Complete, production-ready applications built with 100% local LLMs
Applications:
- 🔬 AI Research-Code Pipeline - Multi-agent system (Researcher → Developer → Validator)
- 🏠 Korean Real Estate AI - Market analysis with government data
- 🤖 Agent Integration Examples - Router patterns, MCP tools & A2A workflows
Multi-Agent Architecture
User Query → Supervisor Agent ↓ ┌───────────┼───────────┐ │ │ │ Research Developer Analysis Agent Agent Agent │ │ │ └───────────┴───────────┘ ↓ Final Result (100% Local)
Intelligent research agent with supervisor-driven architecture
- 💬 Interactive chat with intelligent research triggers
- 🔬 Deep research via GPT-Researcher + local document RAG
- 📚 Multi-format document processing (PDF, DOCX, TXT, MD, JSON)
- 🗄️ Conversation management with SQLite persistence
- 🌐 Hybrid research: Web + Local documents
📦 Simple LLM Chatbot - Cost-effective chatbot with pay-per-use APIs
- Streamlit frontend + FastAPI backend
- Multiple API provider support (OpenAI, Ollama, Anthropic)
- Modular architecture for easy customization
🔍 Simple RAG (No LangChain) - Lightweight RAG implementation
- Pure Python RAG without frameworks
- Support for PDF, PPTX, DOCX uploads
- Semantic search with vector embeddings
🖥️ Local LLM Server Setup - Comprehensive server setup guide
- Ubuntu 22.04 server configuration
- Docker containerization
- GPU support (NVIDIA & AMD)
- Vector database setup
Raw web page → Markdown → Structured data extraction → Neo4j knowledge graph — 100% local LLMs
Here's additional demo you can try on my website.
From $1000+/month cloud bills → ~$50/month electricity (95% savings)
Not just demos - real applications serving real users:
- 2asy.ai - 10 daily automated publications
- Multi-server deployments with 99.9% uptime
- Handling 50+ concurrent users
- Zero data sent to third parties
- Complete control over your infrastructure
- GDPR/compliance ready out of the box
- GPTQ-Marlin kernels: 2-3x faster inference
- Smart model routing across hardware
- Load balancing & automatic failover
- 🚀 Deployed production AI news platform serving 1000+ daily visitors
- 📊 Built monitoring infrastructure for 70+ LLM models
- 🔧 Created reusable MCP services for agent automation
- 🎨 Implemented local vision-language model experiments
- 📰 Developed real-time global news collection system with GDELT