Jeremy Silva

Jeremy Silva is an AI engineer at NVIDIA who previously led product at Freeplay, where he developed tools for agent evaluation, human review, and production monitoring. His work addresses a central challenge of enterprise AI: turning convincing prototypes into reliable products that improve through real-world feedback.

Silva began in data science, building natural-language-processing models for healthcare before GPT-3, then moved into machine learning engineering and infrastructure for scaling models. At Freeplay, he progressed from AI engineering in 2024 to product leadership by the 2025 AI Engineer World’s Fair. His professional profile connects that trajectory to NVIDIA.

His Freeplay work evolved from combining human review with model-graded evaluations into observability for complete agent workflows and automated analysis of production behavior. By 2026, he was describing an agent data flywheel that connects logs, reviewer feedback, evaluations, and experiments into continuous product improvement.

• Evaluate the actual product. Silva favors product-specific evaluation criteria grounded in customer needs, representative examples, and recognizable failure modes. His guide to trustworthy evaluations treats prompts, models, cost, latency, and human judgment as parts of one testing process.

• Turn human review into operational intelligence. Human expertise supplies context and domain knowledge that automated scoring misses. Silva’s work on Review Insights turns reviewer notes and labels into patterns that can inform new evaluations, datasets, and prompt experiments.

• Make AI quality someone’s responsibility. He identifies an emerging AI quality lead who combines customer understanding with systems thinking to define evaluation criteria, investigate failures, label examples, and run experiments. The role can draw from product, operations, engineering, or domain expertise without requiring ownership of production code.

• Build AI into the product strategy. Silva argues against isolated AI initiatives and features designed chiefly to demonstrate technical capability. His approach to integrated AI products starts with customer problems and develops incrementally from embedded assistance toward contextual, interconnected agent workflows.