RELAI

RELAI builds verifiable continual learning infrastructure for teams developing and operating AI agents. Its platform turns failures and human feedback into replayable learning environments, preserving inputs, state, tool calls and memory. Teams can optimize prompts, tools, models, workflows and agent logic against objectives such as task success, latency and cost, then review proposed improvements as pull requests. A learning system of record tracks diagnoses, evaluations and shipped changes.

RELAI was founded in 2024 by Soheil Feizi, a University of Maryland, College Park computer-science associate professor. Its technical approach places regression checks inside optimization: candidate repairs are tested against accumulated learning environments while being developed. The company’s RELAI-VCL research evaluates whether repeated optimization can improve performance on newly introduced tasks while preserving earlier gains, using a two-phase experiment on Terminal-Bench 2.0.

RELAI announced $6.9 million in total funding in 2026 and opened limited-access onboarding; its homepage now advertises a public preview. Its infrastructure works with existing agent frameworks and observability tools, including LangGraph and Braintrust. C3 AI describes using RELAI to turn difficult enterprise use cases into evaluations and measurable improvements in agents shipped to customers.