I build developer tools and AI infrastructure — code-intelligence engines that let AI coding agents start with the codebase already mapped, and the systems work underneath them. Python and C++. 4,600+ GitHub stars across projects people run daily.

Available for senior / staff engineering roles in developer tools, AI infrastructure, and code intelligence. Remote, from Greece (EET — full EU overlap, US mornings). Co-founding conversations welcome as a second track. Reach me on X.

One engine, two front doors.

roam-code — the engine. Run roam index once and an agent starts with the codebase fully mapped: symbols, call graphs, dependencies, architecture and git history in local SQLite. 28 languages, 285 commands, 244 MCP tools, change-safety gates. Fully local, no API keys. pip install roam-code · roam-code.com · docs · source

compile-code — the front door. A thin CLI over the same engine: compile claude indexes, wires and launches in one step, pre-resolving who-calls-this, what-changed and what-breaks before the agent's first token, then verifying the diff after it edits. Measured in its own 41-cell A/B (medians, navigation/comprehension): agent turns 6→1, input tokens −80%, cost −63% — smaller on a second model (−33% turns overall).

• claude-code-guide — the most-starred community reference for Claude Code. Hooks, MCP, settings, skills, the full CLI surface; auto-updated against the official docs every two days.
• super-hermes — skills that teach Hermes Agent (235k★) to write its own analytical prompts, then report both what it found and what it could not see.
• chatbot-injections-exploits — catalog of prompt-injection techniques and exploit patterns. A red-teaming reference since 2023.
• deep-research — fractal multi-agent research. A question spawns angles, angles recurse until atomic, then everything synthesizes back into one answer.

agi-in-md — system prompts are cognitive prisms: they change how a model frames a problem, not how hard it thinks. A 332-word prompt lifts Haiku 4.5 at minimum reasoning to 9.8/10 AI-judged depth on real production code — the prism, not the model tier, carries the depth. 82 prisms, 330+ proven principles, 5,800+ experiments, cross-validated on Claude and Gemini.

AgentsKB — knowledge layer for AI agents: 39,827 verified answers across 244 technologies, returned in ~40 ms, free with no limits. Runs as an MCP server so agents stop guessing about APIs and SDK versions they were never trained on.

• mzip — detection-based compression in C++17. Held-out, type-stratified benchmark over 70 real files: 28 wins, 42 ties, 0 losses. 7.43× overall ratio against brotli-11 at 4.59×, xz-9e at 4.85× and zstd-22 at 4.11×.
• grouped-simd-hashtable — C++ hash table using grouped SIMD metadata scanning. Beats ankerl::unordered_dense, the accepted state of the art, at scale.
• tieredsort — header-only C++17 numeric sorting. 3.6× faster than std::sort, up to 21× on dense data.
• chameleon-cache — variance-aware cache replacement, +5.09pp hit rate over TinyLFU. Developed with feedback from Ben Manes, author of Caffeine.