AI & Backend Engineer in Seoul — I build AI systems that survive contact with production: knowledge graphs, RAG pipelines, AI agents, and the gateways, APIs and UIs that make them usable.
I run sortInc., a small engineering studio. Everything below is public — clone it, run it, read the commits.
- Knowledge graphs & RAG — Neo4j + n10s ontology stores next to pgvector / RAGFlow, so retrieval cites its sources
- LLM gateways & governance — LiteLLM proxies with guardrail chains, PII masking and hash-chained audit logs
- Multimodal vision — a fast detector proposes (YOLO), a VLM verifies (Qwen3-VL)
- Local fine-tuning — LoRA pipelines from synthetic data to GGUF, re-registered in Ollama
- Offline-first stacks — vector-tile maps and digital twins that run with zero runtime network calls
| Project | What it does | Stack |
|---|---|---|
| image-to-ppt | Turns a screenshot, scan or whiteboard photo into a .pptx with native, editable shapes, tables and text — never a flattened image |
Python · FastAPI · React · Vision LLM |
| neo4j-n10s-pipeline | Docker starter for Neo4j + n10s (neosemantics) + APOC — RDF ontology in, a queryable knowledge graph out | Neo4j · RDF · Cypher · Docker |
| llm-finetuning-tutorial | End-to-end LoRA fine-tuning for local Ollama models, text and vision: dataset → train → merge → GGUF → Ollama | transformers · PEFT · Qwen3 / Qwen3-VL |
| llm-chat-widget | Side-panel AI chatbot for Chrome & Edge that answers from the page you're reading, with evidence highlighting | JavaScript · LiteLLM · MV3 |
🌐 sort-tech.github.io/ai-tech-portal · ✉️ sortinc.dev@gmail.com
If something here saved you time, a ⭐ helps others find it.