Ofer Mendelevitch

Ofer Mendelevitch is an AI engineer, startup founder, and technical author focused on making enterprise AI accurate, secure, and accountable. He co-founded the synthetic-healthcare-data company Syntegra, led developer relations at Vectara, and helped develop open-source tools for evaluating retrieval-augmented generation and detecting hallucinations.

Mendelevitch studied computer science at the Technion–Israel Institute of Technology and earned a master’s degree in electrical engineering at Tel Aviv University. His earlier career included engineering leadership at Yahoo and Nor1, an entrepreneur-in-residence position at XSeed Capital, and data-science leadership roles at Hortonworks, LendUp, and Helix. He co-authored Practical Data Science with Hadoop and Spark with Casey Stella and Douglas Eadline.

In 2019, he co-founded Syntegra with physician Michael Lesh and served as chief technology officer. The company applied generative language models to privacy-preserving synthetic healthcare data, creating artificial patient records designed to preserve useful clinical patterns while protecting individual privacy. Mendelevitch and Lesh investigated how to assess both statistical fidelity and privacy; Mendelevitch also addressed the infrastructure demands of training large models, including cloud costs and interrupted computing workloads.

At Vectara, he focused on production-grade retrieval-augmented generation: grounding model outputs in enterprise documents while preserving source attribution, document permissions, and operational oversight. His major technical concerns include:

• Evaluation without golden answers. He contributed to open-rag-eval, which evaluates retrieval and generated responses without exhaustive reference answers or manually labeled passages. Its methods include UMBRELA for retrieval relevance, AutoNuggetizer for answer quality, citation-faithfulness checks, and hallucination detection; connectors support Vectara, LangChain, and LlamaIndex.

• Detecting unsupported model claims. He co-authored FaithBench, a benchmark for difficult hallucinations in generated summaries, and research introducing FaithJudge, which uses human-annotated examples to improve automated assessments of factual faithfulness.

• The operational cost of enterprise RAG. Effective systems must handle document parsing, chunking, hybrid retrieval, reranking, latency, infrastructure costs, continuous evaluation, multilingual support, and document-level access controls. His analysis of enterprise RAG architecture emphasizes that vendor fragmentation and specialized staffing can be as consequential as model accuracy.

• Private-data research and agent coordination. He has described research agents working across internal enterprise data, with applications including RFP responses and employee onboarding.

In 2026, Mendelevitch and Forrest Sheng Bao published Hands-On RAG for Production, covering ingestion, retrieval, evaluation, privacy, security, agentic workflows, multimodal systems, and GraphRAG. Its companion repository provides runnable examples.