Sangin Geum — building tools for local-first AI agents and high-performance C++ systems.
Focused on:
- Agent infrastructure that stays deterministic, inspectable, and fully under the owner's control
- C++20 networking libraries with strong performance characteristics
- Local-first stacks (Ollama + Qdrant) that never require the cloud
| Project | Description | Stack |
|---|---|---|
| code-indexer | Semantic code search for agents. One-shot CLI + MCP server over Qdrant + Ollama. Agents never index themselves — they just search. | Python, tree-sitter, Qdrant, Ollama |
| tackboard | Project-scoped task board for multi-agent coordination. CLI-first, SQLite-backed, dependency DAG, claim system, read-only human board. | Python, SQLite |
| calc-engine | Deterministic, subcommand-driven CLI math engine designed for agents. Strict stdout contract, finance, statistics, distributions, matrices. | Python |
| filekeeper | Guarded, undoable filesystem operations for agents and people: atomic writes, a durable operation journal, and undo that refuses to clobber later changes. | Python |
| Edriel | C++20 multicast pub/sub library with automatic discovery and typed protobuf messaging. Best-effort + reliable QoS paths. | C++20, Asio, gRPC, Protobuf |
- NetworkProgramming — C++ networking examples with Asio
- Algorithm-and-DataStructure — Classic algorithms & data structures in C++
- Various older game / visualization experiments (Gomoku, Breakout, A*, RayCasting, etc.)
- Local AI agent tooling (MCP, CLI-first design, deterministic contracts)
- C++20 systems programming & high-performance networking
- SQLite + vector databases for agent memory and coordination
- Clean, agent-friendly CLI surfaces (one-shot commands, stable JSON, clear exit codes)
Building tools I actually want to use with my own agents.