The big picture: Most AI teams today are “flying blind,” relying on reactive dashboards that only show errors after a user complains. Respan (formerly Keywords AI) is building a proactive observability platform designed to close the loop between evaluation and production, ensuring agents remain reliable at scale.
Why it matters:
- Scale and Growth: Over the last year, Respan has scaled to support 100+ startups and enterprise teams, processing over 1 billion logs and 2 trillion tokens every month for 6.5 million end users.
- Proactive Response: Unlike traditional tracing tools, Respan uses an automated evaluation agent to continuously monitor production behavior, turning drift and hallucinations into actionable prompt updates or alerts before they impact the user.
- Financial Momentum: The company reported over 8x year-over-year revenue growth in 2025, signaling a massive market shift toward “agentic” observability as AI systems become more autonomous.
How it works:
- Observability Gateway: Logs every agent session in production, capturing execution traces, tool calls, and routing decisions.
- Metric-First Evals: Closes the loop from testing to production by continuously scoring output quality based on custom success metrics.
- Prompt Optimization: Includes built-in automatic prompt engineering that allows teams to iterate and deploy updates directly from live production data streams.
The catch: Respan is entering an increasingly crowded “LLM-ops” market where giants like LangSmith and Braintrust already have a head start. While its “proactive action” branding is a strong hook, the challenge will be maintaining low latency while running complex evaluation agents on every production call. Much like the hurdles in agentic security, Respan’s success depends on whether enterprise teams are willing to trust another automated layer to “fix” their non-deterministic software.
Key Details
- Funding: $5M (Seed)
- Lead: Gradient Ventures (Google)
- Founders: Raymond Huang, Andy Li, Hendrix Liu
- Sector: AI Infrastructure / Observability
