Mihai Dragomir

I’m a Senior AI Engineer specializing in LLM-powered systems, retrieval-augmented generation (RAG), and agentic AI architectures. I build and scale production-grade AI platforms end-to-end—from data pipelines and model training to retrieval, orchestration, evaluation, and observability—across AWS and GCP.

At LexisNexis, I led development of a production-grade Legal Summarization Engine on Lexis+ AI, processing long-form case law (100+ pages) into structured headnotes and holdings, reducing attorney review time by 60%+. I also fine-tuned LLaMA models using QLoRA and built hierarchical summarization pipelines for documents exceeding 200K tokens, improving reliability and factual grounding.

I focus heavily on measurement and trust: I built RAG pipelines using Pinecone + Elasticsearch (hybrid search), reduced hallucinated legal claims by ~50%, and delivered a citation validation layer achieving >92% citation accuracy. I established automated evaluation pipelines to track faithfulness, ROUGE-L, and reasoning consistency across model releases, while deploying high-throughput inference with vLLM and TensorRT-LLM for sub-second latency.

Before that, I delivered enterprise conversational AI at Parloa—designing real-time ASR → NLU → Dialogue Manager → TTS pipelines, improving intent recognition by 30%+ and reducing average handling time by 20–25%. At Revolut, I built fraud and transaction anomaly detection models that reduced fraudulent transactions by 30%+ and implemented real-time streaming analytics with sub-second decision latency, alongside experimentation platforms for A/B testing.