Emil R.

I build LLM systems and I define how their quality gets measured, which is a combination most clients have to hire twice for. Most recently I was the AI Expert for the Government of El Salvador, where I architected a cross-institutional document-validation API on FastAPI and Claude serving four government institutions. It confines the LLM to structured JSON extraction and keeps deterministic rules at the decision core, so every outcome is auditable. A template-driven rules engine meant a new document type could be onboarded with zero lines of code, through a single JSON template defining fields, rules and confidence thresholds, with 34 document types targeted for the Ministry of Environment. I grew that codebase from 12 to 128+ tests and migrated inference to Claude on Vertex AI with application-default credentials, removing every API-key secret from the system. At Turing I worked as AI Engineer and Post-Training Data Lead, co-developing five specialized agents inside a team of seven that handled 500+ weekly queries at 20% lower latency, built with LangGraph DeepAgents, LangChain and Microsoft Agent Framework over OpenAI, Azure OpenAI and Anthropic, containerized with Docker on Azure. The other half of that role was post-training data: reward structures for agentic RL, Chain-of-Thought authoring for Python and logic domains, and Python eval harnesses for RLVR that replaced manual prompting with programmatic scoring. The evaluation side goes back almost a decade. At Magic Data I delivered 800+ multi-turn evaluation cases and 150+ side-by-side model comparisons at a 96%+ QA acceptance rate. I still work with Alignerr, Outlier and Appen on model evaluation and relevance assessment, and I run the final Swedish-language approval gate on roughly 100 AI-dubbed videos a week against a 10-criteria client rubric. On Upwork I held Top Rated Plus with a 100% job completion rate across 8+ years. What I bring to the language side is six working languages and an M.A. in Linguistics from Waseda University in Tokyo, plus 13 years as the public Swedish-language expert on WordReference, where I defended usage calls to everyone from beginners to academic linguists. Where I can help: Agentic AI: LangGraph, LangChain, CrewAI, Google ADK, Microsoft Agent Framework, MCP servers, tool calling, RAG over vector databases Production Python: FastAPI, Pydantic, pytest and BDD, uv and ruff, Git, Docker, CI/CD Cloud: AWS (Lambda, S3, SageMaker, API Gateway, EKS), Google Vertex AI, Azure, LangSmith tracing and observability Evaluation and data: rubric design, RLHF and SFT preference data, inter-annotator agreement, multilingual annotation in English, Swedish, Japanese, German and Spanish Open to AI engineering, agent development, LLM integration and evaluation work. Three changes worth knowing about. Your current text says you currently work with Turing and Magic Data, but both of those ended (Turing in January, Magic Data in November last year), so I moved them to past tense and led with the government role instead. Your degree is an M.A. in Linguistics with a Creative Writing and Discourse Analysis specialization rather than an M.A. in Creative Writing, and the precise version is more impressive for AI work anyway. And I left out the decade across Tokyo, Berlin and Beijing, because I have Tokyo confirmed but nothing on Berlin or Beijing; if those are real, tell me the years and I’ll put a tightened version back in. One line I deliberately cut: “systems that don’t just speak; instead, they reason, integrate, and act.” It’s the kind of phrasing clients now read as AI-written, which is an awkward signal on an AI engineer’s profile.