What I'm building — a flagship product and a suite of agent, harness, and MCP tooling.
An ultra-reliable Agentic-RAG workspace that turns fragmented corporate data into actionable intelligence — surfacing a living 3D “Singularity” of your organization's knowledge graph, guided by the AI persona Carthor. Multi-platform Flutter (Web PWA · macOS · Windows · Linux) over a Java agentic-RAG backend.
A multi-tenant agent platform where organizations evolve agents with auditable confidence: every promoted skill must beat the previous baseline on replayed historical trajectories before it ships.
The engineering base for harness-based agent operation. It injects the right knowledge (RAG) into an AI session while exposing execution tools over MCP — separating what the agent knows from what it can do, and keeping the context window lean.
An async microservice that reacts to external events and autonomously generates code by running the Claude Code CLI inside sandboxed workspaces — every step captured in a structured trace, results delivered by callback.
A standalone MCP server that lets LLM agents compose rich, interactive Slack messages with Block Kit — stateless, with no Slack tokens or network access required.
An interactive playground where junior developers practice and compare RAG, prompt engineering, LLM parameters, agent tool-calling, and GraphRAG hands-on. Spring AI backend, Next.js front, Neo4j + Qdrant.
A generator for low-poly animations and wireframe visuals for the web — the same generative-geometry spirit that drives this site's hero.
I don't chase technology trends — I work to understand a technology's essence and limits, then convert it into value for the organization and its customers. Across sixteen years a consistent way of working emerged: robust design and strong execution, with agents, knowledge, and evidence fitting together.
I build the operating layer before the feature. A reusable harness injects knowledge (RAG) and exposes tools (MCP), so an agent does real work on a lean context — what it knows and what it can do, kept cleanly apart.
No skill ships on a hunch. Every change must beat the previous baseline on replayed history — improvement you can measure and audit, not conviction without proof.
I treat the agent as a builder. Specs become running code through autonomous, event-driven execution — prototyping at conversational speed, with every step traced and reproducible.
A model's output is only as good as what it's given. I design what the model sees — retrieval, memory, and tool surfaces — so the right knowledge arrives at the right moment.
Design documents don't ship products. I stay hands-on — writing code when it matters, digging into performance issues and failure root causes — and hold responsibility until the project becomes a real result.
From founding a startup to national R&D, I weigh technical completeness against business viability — validating new models fast without mortgaging the traffic, data, and organization the service will need after it grows.
Building something at the intersection of AI, scale, and good engineering? I'd love to hear about it — reach me on LinkedIn or by email.
Neil Kwon · neil@excogitar.com
