Rayehe H.

I build LLM agent systems that hold up in production — LangGraph orchestration, RAG over messy real-world data, and the evaluation layers that keep them reliable. Recent work: a multi-agent platform managing 13 interdependent workstreams with automatic dependency invalidation and a reviewer agent gating output; a retrieval layer combining keyword and semantic search with cross-encoder reranking and MCP tool calling against live APIs; a LangGraph conversational agent with deterministic guardrails living outside the model and an adversarial evaluation harness; a healthcare claims pipeline combining deterministic rules with fine-tuned models under HIPAA, where auditability decided the architecture. Also a full-duplex voice assistant over an ERP backend — streaming ASR, semantic endpointing, barge-in, WebRTC. Most of my time goes to routing, evaluation and failure handling rather than model calls. Integrating an LLM API takes a week; making it reliable enough that someone acts on the output is the job. Stack: Python, LangGraph, FastAPI, PostgreSQL + pgvector, Docker, Anthropic/OpenAI/Gemini, self-hosted inference via vLLM and Ollama. 7+ years, MSc in Artificial Intelligence, two published papers (ACL, SemEval).