Respan is a proactive, full-stack AI engineering platform that provides a unified control plane for developers to trace, evaluate, and monitor LLM agent behavior while offering an AI gateway to improve production performance. Pulse 2.0 interviewed Respan founder and CEO Andy Li to learn more.
Andy Li’s Background
Could you tell me more about your background? Li said:
“I’m Andy Li, co-founder and CEO of Respan. I studied mechanical engineering and economics at UIUC, where I met my co-founders and some of my closest friends. Before starting the company, I worked at Apple as a product design engineer on parts of AirPods 4.”
“I left school to build full-time, and we joined YC W24 shortly after. We originally started as Keywords AI, building infrastructure for dynamic LLM routing. But through conversations with customers during YC, we realized the bigger problem was not routing itself. AI teams were starting to ship complex agents into production, but they had almost no visibility into how those agents behaved, why they failed, or how to improve them over time.”
“That insight led us to rebuild the company around Respan: proactive observability and evaluations for AI agents. Today, we help teams trace, evaluate, and improve agent behavior from development to production. We’re based in the Bay Area, work out of hacker houses in Alameda, and now serve more than 100 customers.”
Formation Of The Company
How did the idea for the company come together? Li shared:
“The idea came from our own experience building AI infrastructure. We started by building an API for dynamic LLM routing, helping developers route requests across OpenAI, Anthropic, and other model providers. But during YC, we kept hearing the same pain from AI companies: once agents were in production, teams had very little visibility into what was happening.”
“They could see that something broke, quality dropped, or costs spiked, but they often could not explain why. The existing tooling was mostly built for single LLM calls or traditional observability. It was not built for multi-step agents that plan, call tools, use memory, and make decisions over time.”
“That became the core insight behind Respan. Teams did not just need another routing layer. They needed a system to understand, evaluate, and improve agent behavior continuously in production.”
“Today, Respan helps teams trace, evaluate, and optimize AI agents from development to production. The goal is to close the loop: not just showing teams what broke, but helping them fix it faster.”
“As CEO, I own product direction, customer conversations, hiring, and fundraising. I stay very close to product and design because at this stage, company strategy and product direction are basically the same thing.”
Favorite Memory
What has been your favorite memory working for the company so far? Li reflected:
“My favorite memory is working and living with my core founding team in a hacker house together in the Bay Area. We started the company as close friends from UIUC, and moving into the same house made the company feel much more real. Everything became faster: product decisions, customer calls, late-night debugging, hiring conversations, and the random ideas that only come up when you are around each other all the time.”
“It also shaped the culture of Respan. We are building in a very intense, high-trust environment, and the hacker house made that real from day one. Some of my favorite moments have been the ordinary ones: sitting around after a long day, arguing about product direction, fixing something together, or realizing at 2 a.m. that we had just solved a problem customers really cared about.”
Core Products
What are the company’s core products and features? Li explained:
“Respan is an AI observability and evaluation platform for teams building AI agents. We help companies trace, evaluate, and improve agent behavior across development and production.”
“Our core product starts with observability. Teams use Respan to see what their agents are actually doing in production through traces, logs, metrics, dashboards, and alerts. This helps them understand failures, debug complex workflows, and track cost, latency, and quality over time.”
“The second layer is evaluations. Respan helps teams build datasets, run offline and online evals, compare prompts and models, and catch regressions before they reach users. This is especially important for agents, where behavior can change across multi-step workflows and is harder to evaluate with traditional tests.”
“We also provide prompt optimization workflows, so teams can test, version, and improve prompts based on real production behavior and evaluation results.”
“Finally, we have an AI gateway that handles the infrastructure layer: routing requests across model providers, managing fallbacks, caching, retries, and monitoring usage across providers.”
“The broader goal is to bring these pieces together into one loop: trace what happens in production, evaluate the behavior, improve the system, and deploy changes with more confidence.”
Challenges Faced
Have you faced any challenges in your sector of work recently? Li acknowledged:
“One of the biggest challenges is that the AI infrastructure market is changing extremely fast. A year ago, most teams were experimenting with simple LLM workflows. Now many are shipping multi-step agents into production, where failures are much harder to understand. An agent can call tools, use multiple models, retry, hallucinate, fail silently, or create unexpected downstream behavior. Traditional software observability was not built for that.”
“We handled this by staying very close to customers and rebuilding around real production problems. Instead of building another generic dashboard, we focused on the full loop: tracing agent behavior, evaluating production data, monitoring regressions, and helping teams move from ‘something broke’ to ‘here is what happened and how to fix it.’”
“The other major challenge is trust. AI infrastructure often touches sensitive production data, so customers need to know the platform is secure and reliable. We have invested heavily in security, compliance, and better admin controls, so teams can get the visibility they need while maintaining strong data governance and control.”
Evolution Of The Company’s Technology
How has the company’s technology evolved since launching? Li noted:
“We started as an API for dynamic LLM routing. At the time, the problem we were focused on was helping developers choose the right model based on cost, latency, and quality.”
“As customers started putting AI systems into production, we realized routing was only one part of the problem. The bigger need was visibility. Teams needed to understand what was happening inside complex AI workflows: why agents failed, which steps caused issues, how prompts and models performed, and whether changes were actually improving quality over time.”
“That led us to expand from routing into a full AI observability and evaluation platform. Today, Respan includes production tracing, logs, metrics, dashboards, alerts, automated evaluations, prompt experiments, dataset workflows, and gateway infrastructure.”
“The biggest shift is that we moved from helping teams send AI requests to helping them understand and improve the full lifecycle of their AI agents.”
Significant Milestones
What have been some of the company’s most significant milestones? Li cited:
“Some of the biggest milestones have been getting into YC, rebuilding the company around AI observability, rebranding from Keywords AI to Respan, raising our seed round, and growing into a platform used by more than 100 AI teams.”
“YC W24 was a major turning point. It forced us to move fast, talk to customers constantly, and be honest about the problem we were actually solving. That process led us to move beyond dynamic LLM routing and focus on observability and evaluations for AI agents.”
“The rebrand to Respan was another important milestone because it reflected where the company was really going. We were no longer just building infrastructure around LLM requests. We were building the observability and improvement layer for production AI agents.”
“More recently, raising our $5 million seed round and scaling the platform to process billions of logs and trillions of tokens per month has been a big step. It gives us a unique view into how AI systems are breaking in production and what the next generation of AI infrastructure needs to look like.”
Customer Success Stories
Can you share any specific customer success stories? Li highlighted:
“One strong example is Retell AI, which helps companies build human-like phone agents. Their systems operate at a very high scale, with millions of LLM calls across thousands of concurrent conversations. For a company like Retell, observability is not a nice-to-have; it is a core operational requirement.”
“Their CTO said, ‘We scaled from 5M to 500M+ monthly API calls quickly. Respan gave us the debugging layer to resolve production issues 10x faster.’”
“At that scale, every log needs to tie back to a specific call or conversation, and the team needs to debug failures quickly without losing production visibility. Retell tried multiple tools, but Respan was the one that could reliably support their volume, process production logs, and help their team maintain quality as usage scaled.”
Funding/Revenue
Are you able to discuss funding and/or revenue metrics? Li revealed:
“Respan recently raised a $5 million seed round led by Gradient, with participation from Y Combinator, Hat-Trick Capital, XIAOXIAO FUND, Antigravity Capital, Alpen Capital, and a number of founders and operators across the AI ecosystem.”
“On the traction side, we are trusted by more than 100 AI startups and enterprise teams. The platform processes more than 1 billion logs and 2 trillion tokens per month, supporting over 6.5 million end users. We also grew revenue 8x year-over-year in 2025.”
Total Addressable Market (TAM)
What total addressable market (TAM) size is the company pursuing? Li assessed:
“We see this as a very large market because AI observability sits at the intersection of application performance monitoring, developer infrastructure, AI evaluations, and gateway infrastructure.”
“Every company shipping AI agents needs to understand what those agents are doing in production: how they perform, where they fail, how much they cost, and how to improve them. That need only becomes more important as software shifts from deterministic workflows to AI-driven agents.”
“Our view is that observability and evaluation will become mandatory infrastructure for AI-native software, the same way monitoring became mandatory for cloud software. If agents become core to enterprise workflows, the market for understanding and improving those agents becomes a multi-billion-dollar category.”
Differentiation From The Competition
What differentiates the company from its competition? Li affirmed:
“Most observability tools are still dashboard-first. They surface logs, traces, and metrics, but they often stop at showing what happened. For AI agents, that is not enough. Engineers still have to manually connect production traces, evaluation results, prompt changes, and model behavior to understand why something failed.”
“Respan is built around closing that loop. We combine observability, evaluations, prompt optimization, and gateway infrastructure in one platform so teams can trace what happened, measure quality, detect regressions, and improve the system faster.”
“The biggest difference is that we are metric-first and evaluation-driven. We do not want to just give teams more logs to look at. We want to help them understand how their agents are performing in production and give them a path to make those agents more reliable, efficient, and accurate over time.”
Future Company Goals
What are some of the company’s future goals? Li emphasized:
“In the near term, we are focused on scaling the team and continuing to invest in infrastructure as customer usage grows. We are processing a large and increasing volume of production AI traffic, so reliability, performance, and security are major priorities for us.”
“On the product side, we are focused on helping AI teams move from reactive debugging to proactive production monitoring. That means better tracing, stronger evaluations, faster prompt optimization, and more automation around detecting and fixing agent failures.”
“The bigger vision is self-driving observability for AI agents. We want Respan to not only show teams what is happening in production, but also identify issues, generate evaluations, suggest fixes, and prevent regressions before they reach users.”
“Long term, our goal is to become the default observability and improvement layer for every team shipping AI agents in production.”
Additional Thoughts
Any other topics you would like to discuss? Li concluded:
“AI infrastructure is entering a new phase. Most companies have moved past the prototype stage and are now shipping real agents into production. But the tooling has not kept up. The hard part is no longer just calling an LLM. It is understanding how agents behave in production, why they fail, and how to improve them continuously.”
“That is the infrastructure gap Respan is focused on. We believe observability and evaluation need to be part of the core development workflow, not something teams bolt on after production issues happen.”
“The other topic is team culture. We are a small, intense team working out of Alameda, and we care deeply about speed, customer proximity, and building directly with the teams pushing AI into production. That operating style is a big part of why we have been able to move quickly.”
