AI/LLM engineer. Python and TypeScript. From Yakutia (−60°C winters), now in Vietnam (UTC+7), working remote.
I build LLM-era backends — retrieval, agents, MCP servers — and I like finishing things properly: tests on the parts that can hurt, honest READMEs, post-mortems when something fails.
galtonian-context-retrieval — research: RAG retrieval as a stochastic particle process over a semantic graph, with an adaptive μ+kσ context threshold. Benchmarked honestly (dense top-k wins on the small corpus; adaptive context sizing holds — 40% smaller context at unchanged recall). Paper draft, multi-seed runs, bootstrap CIs.
ai-ops-hub — MCP server that gives AI clients sandboxed hands on a machine: files, tasks, web, and hybrid search (SQLite FTS5 + embeddings, rank fusion) over a personal corpus. The security paths — path traversal, host allowlists — are the tested part: 45 unit tests, CI.
Game-master-v3 — an AI game master with a persistent world. Async FastAPI over four stores doing four jobs: Neo4j (knowledge graph), Qdrant (semantic search), PostgreSQL (event sourcing with real snapshot rollback via reverse replay), Redis (cache). D&D 5e dice engine, procedural worldgen. mypy strict, 108 tests that run without any of the databases.
Freelance web development (landing pages, analytics, automation). Web3: Solana/Anchor programs, a Cosmos SDK chain, and a win at the ICP Hackathon Hong Kong 2024 (Bridge23). Before software — control systems for nuclear power plants (China, Egypt) and international logistics. The common thread: systems where mistakes are expensive.
Memory mechanisms LLMs don't have: episodic event graphs, recognizing individuals rather than categories, forgetting as a feature. A long-running private project — multi-seed test protocols, pre-registered design decisions, written post-mortems. Ask me about it.