Anton C.

I'm a backend engineer. For the last two years I've worked on a legal AI product for defense attorneys, where a hallucinated citation isn't a funny screenshot, it's someone's case. Most of my time goes to the boring parts nobody demos: why retrieval returns the right document ranked fourth, why an agent loops on a tool call, why the thing that passed review in March quietly got worse by July. That last one is the hard one. LLM systems don't crash when they degrade, they just start being a bit wrong, and you won't notice unless you built something that measures it. Concretely, I work with: Multi-agent orchestration — tool calling, SSE streaming, routing across Groq/Gemini/Anthropic/OpenAI to keep cost and latency sane. Search. This is the part I actually enjoy. Deep OpenSearch work: hybrid vector + BM25, nested and reverse_nested aggregations, custom script_score ranking, hierarchical documents where a paragraph only makes sense inside its article. Qdrant when a dedicated vector store fits better. Most RAG problems are search problems wearing a costume. Async Python that holds up — FastAPI, SQLAlchemy 2.0 async, PostgreSQL, Dramatiq/Redis, Docker. If your AI works in the demo and falls apart on real documents, I've probably already met that bug.