Research Engineer · WeDataLab AI Research Lab · Seoul
Hi, I'm Jaesol.
I build AI infrastructure — observability, retrieval, and the database layer underneath. I work mostly in Python and PostgreSQL, and I write about agents, inference, and the systems they run on. I am interested in where LLM agents meet social complex systems, and I study computational neuroscience with the ICARUS research group.
- Focus
- AI infrastructure, observability, RAG
- Works in
- Python, PostgreSQL, Go
- Writes about
- Agents, Databases, Inference
01 Current work
Research Engineer, WeDataLab AI Research Lab
- Developing AI agent applications
- Developing MCP (Model Context Protocol) servers
- Building RAG pipelines and an LLM observability stack
- Automating infrastructure monitoring
- Researching the adoption of GPU databases
- AI adoption consulting and proposal writing
- Consulting on knowledge base design for research institutions
- Lectures on AI-assisted development
02 Projects
PostgreSQL extensions
- pg_cuvsfeaturedC↗
pg_cuvs adds GPU vector search to PostgreSQL without changing a single query. Selected for a talk at PGConf.Asia 2026 in Hong Kong. Built on pgvector and NVIDIA cuVS, it lets the GPU generate candidates while planning, MVCC, filters and transactions stay in PostgreSQL. If serving should stay on the CPU, the GPU can do the index build alone and hand back a standard pgvector HNSW index, so the GPUs already running your embedding models can take on the slow rebuilds.
- pg_aidbIn-database RAG done right: one-line ai.ask() in SQL, zero blocked connections.C
- pg_glotBeyond naive tsvector: opt-in tokenizers, BM25, and hybrid retrieval for PostgreSQL.Rust
- pg-extension-labA harness for testing, benchmarking, and tuning PostgreSQL extensions in an isolated, reproducible environment.Python
Agent tooling
- cctracefeaturedGo · Next.js↗
cctrace shows what your coding agents actually did. What a session cost, which project burns the most tokens, what the agent was doing at 3pm: telemetry alone can't answer these, so cctrace joins OTLP usage and cost from Claude Code and Codex CLI with the session conversations and tool calls. It is self-hosted: a Go server stores everything in TimescaleDB and serves the dashboard, and nothing leaves your machines.
- xsmWith xsm alone, the Claude Code and Codex sessions you already run become a multi-agent system. No separate runtime, no orchestrator, no wrapper: all it takes is a message channel, and the agents do the rest. Sessions find each other, split the work, hand it off and review each other, and who does what is settled in conversation. What xsm builds instead is communication you can rely on: a message wakes an idle session, every result says how far it got, and an opt-in P2P network reaches your other machines.Python
- journey-qaPersona-journey acceptance QA skill: clean-room journeys, leak-gated evidence and docs screenshots.Python
- jstackWorkflows, skills, and hooks for development, research, and document work in Claude Code and Codex.TypeScript
03 Upstream contributions
- LightRAG featured Major contributor HKUDS/LightRAG ↗
One of the major contributors to LightRAG, the graph-based RAG framework with 40,000 GitHub stars: top 10 by commits, with seven merged pull requests. I built PGTableGraphStorage, a graph backend on plain PostgreSQL tables with no Apache AGE, so LightRAG runs on managed PostgreSQL where AGE can't be installed. I also added an offline tool that moves an AGE graph across without re-indexing documents, and blocked the path where AGE 1.8 crashes the PostgreSQL backend.
PGTableGraphStorage (#3103)AGE migration tool (#3620)AGE 1.8 crash guard (#3621)All commits
- oMLX
- WeKnora
- Senpi
- Graphify
- OpenDataLoader
- Unsloth
04 Writing
- featured preprint ↗
Optimization and Discovery in AI Scientist Systems: Tracing Evidence, Claims, and Decisions
Zenodo · revised and expanded from my CAISc 2026 paper
Better performance alone does not show that an AI Scientist system made a discovery. This study separates optimization from discovery: which claims a result supports, and why the next experiment is chosen. It reviews 21 published configurations with POMDPs and Bayesian experimental design, and checks five systems (ScientistTwo, Adam, PiEvo, Coscientist, Robin) against their sources to tell verified results from interpretation. It offers a way to evaluate them as weighing evidence under uncertainty, not as optimizers.
Zenodo (DOI 10.5281/zenodo.23633018)CAISc 2026 version (OpenReview)
- xsm: A Message Channel for Agent Collaboration
- Weight-Space Programming
- Minority Responses in Persona-Prompted LLMs: Following Architect of Silence
- How Managed Agents Takes an Agent Apart: Brain, Hands, and Session
05 Bio · Ideas
I design enterprise AI infrastructure and study the intersection of LLM agents and social complex systems.
Since attending ALife Kyoto 2025, I have been interested in modeling the collective behavior of AI agents from the perspectives of Computational Sociology and Cultural Evolution.
Ideas
- Modeling the exteriority and systematicity of The Social with agent-based models (ABM)
- Simulating how groups of LLM agents form and propagate social norms
- Formal extensions of Collective Predictive Coding and their application to social theory
- Self-evolving music using Cellular Automata
06 Certifications · Education
Certifications
- Jul 2025 Google Professional Machine Learning Engineer
- Nov 2024 AWS Certified Cloud Practitioner
- Dec 2023 Engineer Big Data Analysis (Korean national technical qualification)
- Oct 2023 SQL Developer (SQLD)
Education
- 2019–2021 Ritsumeikan University, Graduate School of Sociology, M.A. (graduated)
- 2018–2020 Chung-Ang University — Graduate coursework in Sociology
- 2012–2017 Korea National University of Arts (K-Arts), School of Film, TV & Multimedia, Department of Cinema Studies, undergraduate
- 2012 Korea Science Academy of KAIST, graduated
07 Contacts
- GitHubgithub.com/ysys143
- Emailysys143@naver.com