Python backend engineer working on AI/LLM systems. Vietnam (UTC+7), remote.
I build the infrastructure around models rather than the models themselves: agent orchestration, memory and retrieval, tool calling, async pipelines, and the boring parts that decide whether any of it survives production. Currently building a multi-agent research platform at NUAH/Universa.
I care about what happens when a system fails, not only when the happy path works.
galtonian-context-retrieval — retrieval research: RAG as a stochastic particle process over a semantic graph, with an adaptive μ+kσ threshold instead of static top-k. 40% smaller context at unchanged recall@20. Multi-seed runs, bootstrap confidence intervals, ablations, paper draft. It also publishes a negative result: plain dense retrieval beat my method on the benchmark corpus, and the paper says so, because that is what the data showed.
ai-ops-hub — MCP server giving AI clients sandboxed access to files, tasks, web and hybrid search over a personal corpus (SQLite FTS5 and embeddings fused with Reciprocal Rank Fusion). One tool registry behind two transports. The security paths are the tested part: path traversal, host allowlists, 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 event replay, Redis cache. LLM output validated against world state, Prometheus metrics, mypy strict, 108 tests that run without any live database.
Python (FastAPI, asyncio, SQLAlchemy, pytest) · PostgreSQL, pgvector, Qdrant, Neo4j, Redis · RAG, hybrid search, reranking, AI agents, tool calling, MCP · Kafka, Redis Streams, queue workers · Docker, AWS, CI/CD, Prometheus + Grafana · TypeScript/Node · Go (Cosmos SDK), Rust (Solana/Anchor)
Two years building AI systems commercially: a multi-agent research platform (memory architecture, RAG pipelines, MCP tools, cost-aware model routing, 400+ tests), an AI platform for corporate communications at Bridge23, and the backend of a FinTech/RegTech/Web3 platform I designed and wrote on my own.
Web3 along the way: six Solana programs in Rust/Anchor, Cosmos SDK modules in Go, ICP canisters. Our team won the ICP Hackathon Hong Kong 2024.
Before software, six years designing safety-critical automation for nuclear power plants at Rosatom, projects in Egypt and China. Automatic response logic, multi-level verification, GOST and IAEA compliance. That is where the interest in irreversible operations and cost of failure comes from, and it never went away.
Memory mechanisms LLMs don't have: episodic event graphs, recognizing individuals rather than categories, forgetting as a feature. A long-running private project with multi-seed test protocols, pre-registered design decisions and written post-mortems. Ask me about it.