Sanjar Ğabithan

Built a production retrieval-augmented memory service for an LLM agent from scratch: hybrid retrieval (BM25 plus vector cosine, fused with Reciprocal Rank Fusion), LLM-driven structured extraction with rule-based fallback, contradiction handling over an evolving knowledge store, and graceful degradation when the model backend is unavailable, all containerized behind a FastAPI service. I care about the parts that make AI systems survive contact with production: latency budgets, failure modes, retrieval quality, and clean service contracts. My foundation is real-time ML systems on real hardware (KAUST MSc, GPA 3.88), where I cut a perception pipeline's per-frame latency from minutes to milliseconds, and multi-model AI orchestration, including a 1st-place hackathon pipeline that gated heavy foundation models and a GPT-4o reasoning layer behind a lightweight filter for cost control. I am looking to apply this systems-and-orchestration mindset to enterprise LLM infrastructure and to grow into an architecture-owning role. Native Qazaq and Russian, professional English, based in Astana.