I build ontology engines and GraphRAG search, and agent harnesses that own judging, retries and memory.
Then I ship them into banking, capital and retail customers' on-prem environments and measure them there.
- Ontology engine & search systems: LLM-free ontology builds from documents, tables and databases, a PostgreSQL-native graph store, and one-shot GraphRAG retrieval with provenance.
- Agent harness research & development: a state-machine runtime that owns judging, retries, termination, memory and tool search, compiled into PyPI · npm · MCP packages.
- XGEN maintenance & enterprise delivery: porting the platform into air-gapped and on-prem customer environments, fixing what breaks there, and re-measuring under the same conditions.
| Area | What I can do |
|---|---|
| Ontology & Retrieval | Ontology modeling and build pipelines (document / CSV / RDB), incremental sync with PK keyset and row fingerprints, graph storage on PostgreSQL, GraphRAG fusion, hybrid ranking (BM25F · PPR · MMR), evidence and citation tracking |
| Agent Engineering | Agent runtime design, judge · retry · termination policy, agent memory, context engineering and progressive tool disclosure, MCP servers, HITL approval and OAuth 2.1 PKCE |
| LLM Evals | Fixed gold sets, LLM-as-a-Judge, controlled model and agent comparisons, verdict determinism, RAGAS |
| Enterprise Delivery | Air-gapped k3s · Helm deployments, PII masking · guardrails · audit trails, vLLM serving, dependency and SBOM review, Next.js · FastAPI product work |
| Data Science | Statistics, time-series forecasting (DLinear · TimeMixer ensembles), ML prediction and feature importance, clustering, text mining |
| Project | What it does | Proof | Links |
|---|---|---|---|
| xgen-omnifuse | Backend-agnostic one-shot GraphRAG search that fuses vector, graph and HippoRAG expansion into a single synthesis | Public financial-regulation benchmark, multi-hop strict answers 19 → 103 / 120 with no LLM in retrieval |  |
| xgen-ontology-build | Documents, tables and DB rows into a provenance-tracked ontology, LLM-free by default | Output diff 0 against the product pipeline on a 14,924-chunk real corpus |  |
| xgen-harness | 10-stage state-machine agent execution engine; compiles canvas agents into PyPI · npm · MCP packages | Under the same strict-citation judge, scores that split between 0 and full marks settled to 0.70 – 0.98 |  |
| Jermes | Self-improving agent that promotes a procedure to a skill only when its gain reproduces on held-out runs | Cumulative accuracy 44.2% → 88.9% over 190 user questions |  |
| XGEN MAKER | Development-automation harness that lands natural-language requests on exact files and lines through a code ontology of 23 repos | R@10 0.837 on 2,827 human-merged MRs | |
| Agethos | Personality, memory and emotion for LLM agents, with a draft · judge · repair persona compiler | Persona compile fidelity 0.930 |  |
| Year | Award | Conference |
|---|---|---|
| 2025 | Excellence Paper Award | Korea Society of Electronic Commerce & Smart Media Society |
| 2024 | Excellence Paper Award | Korea Society of IT Services |
| 2024 | Best Paper Award | Korea Intelligent Information Systems Society |
| 2024 | Excellence Paper Award | Korean Academy of Management |
| 2023 | Excellence Paper Award | Korea Society of Information Systems |
| 2023 | Best Paper -- Honorable Mention | KHU Big Data Graduate Student Conference |
| 2023 | Grand Prize | Korea Knowledge Management Society -- Idea Competition |
| Experience | Plateer AI-Lab · AI Engineer / FDE (current) Ontology engine & search systems · agent harness research & development · XGEN maintenance & enterprise delivery Shaveron · Freelance Developer (FDE), dispatched to LG CNS Singlex Strategy/Operations Team RAG chatbot prompts, retrieval and evaluation automation |
| Education | M.S. Big Data Applied Science, Kyung Hee University B.S. Statistics, Jeonbuk National University |
| Interests | Agentic AI, Ontology & GraphRAG, LLM Evals, Multi-Agent Systems, Time-Series Forecasting |
Languages
AI & Data
Backend & Frontend
DevOps & Infra
