Nurgissa Sultanmakhmutuly

Built and shipped a production multi-tenant WhatsApp assistant serving three businesses from a single Flask process: RAG over a ChromaDB knowledge base (OpenAI embeddings, hot-reloadable index), GPT-4o-mini with tool calling, and auto-detected bilingual RU/KK dialogue with sticky per-flow localization. Designed a five-agent LangGraph question-generation engine using heterogeneous model routing: deterministic Python for mathematical integrity, low-cost models for generation, reasoning models for a Critic reflection loop with bounded backward retries, and Claude Sonnet for student-facing diagnostics. Applied practical LLM cost engineering: batch pre-generation, prompt caching, and routing each graph node to the cheapest model capable of the task. Strong mathematical background with a focus on applied NLP and scalable LLM systems.