Nir Gazit

Nir Gazit co-founded Traceloop and co-created OpenLLMetry, an open-source framework for observing language-model applications using OpenTelemetry. His work gives engineers practical ways to inspect AI workflows, diagnose failures, and measure whether changes actually improve production systems.

Gazit earned a master’s degree in computer science from the Hebrew University of Jerusalem. At Google, he was a technical lead and worked on machine-learning models supporting Photos, Maps, and YouTube. He subsequently became chief architect at Fiverr and contributed to OpenTelemetry.

In 2022, Gazit and Gal Kleinman began experimenting with language models and found that AI applications lacked the testing and monitoring infrastructure expected of conventional software. They founded Traceloop, with Gazit as chief executive and Kleinman as chief technology officer.

In October 2023, Gazit introduced OpenLLMetry, extending OpenTelemetry to capture model calls, prompts, vector-database activity, and application traces. The open-source project added instrumentation for model providers, retrieval systems, and application frameworks; its semantic conventions became part of OpenTelemetry.

Traceloop developed that foundation into a commercial platform for tracing, evaluation, hallucination detection, and production monitoring. Gazit announced a $6.1 million seed round in May 2025. In March 2026, he announced that Traceloop was joining ServiceNow, integrating its technology into ServiceNow AI Control Tower while keeping OpenLLMetry open source.

How Gazit approaches AI reliability

• Vendor-neutral AI observability. OpenLLMetry instruments model providers, vector databases, and orchestration frameworks using standard telemetry that teams can send to their existing monitoring platforms. Gazit emphasizes practical diagnostics—including retrieved documents, similarity scores, latency, and model calls—and collectors that filter sensitive information before transmission.

• Evaluation-driven prompt optimization. Gazit replaces intuitive prompt tinkering with datasets specifying what successful answers must contain. In his automated prompt-improvement demonstration, a model-based evaluator scores responses, while an agent analyzes failures, proposes revisions, and tests them again.

• Separate retrieval failures from generation failures. A chatbot’s incorrect response may originate in poor document retrieval or in a model’s mishandling of useful context. Gazit evaluates individual components alongside complete workflows, making failures easier to locate and correct.

• Test the tests. Model-based judges can produce inconsistent scores, and repeatedly optimizing against identical examples creates overfitting. Gazit advocates stronger validation through deterministic metrics, statistical safeguards, and separate training and test datasets.

• Monitor model and data drift. AI applications can deteriorate when model behavior, incoming data, or retrieved documents change without a corresponding code change. Gazit pairs production traces with ongoing checks for relevance, faithfulness, hallucinations, and regressions.