The router was the pitch. The logs turned out to be the company.

Fifty characters was all they got, and they spent it on a joke: "API Bartender for LLM APIs."

The product underneath was a router. An ML model sorted each task by complexity and specialisation, then sent it to whichever LLM was best in that category. Cheaper models did the work they could handle. The application states the margin as a fact of the moment: "Currently, we are 35% cheaper than GPT-4 without sacrificing quality."

Then it makes a prediction in a place applications usually make excuses. Asked who the competitors are and who they fear most, the answer is not a moat:

"There are no companies working on a combined LLM as of now, and we firmly believe this is the future for LLM usage."

Both halves of that sentence came true. Almost nobody was doing it in late 2023. It did become the future: routing layers like OpenRouter and LiteLLM are ordinary furniture now, and a written routing policy is the sort of thing founders keep in the repo next to the linter config.

The company is not called Keywords AI any more.

What they had when they wrote it

Not much. The application does not pretend otherwise.

The public beta had launched the day before the progress section was written, on 12 October. Seventeen people were running a free trial key. Seven B2B customers were paying, and the founders admit they cannot yet say how much: "we expect ~$500-1500 /user per month based on their current AI usage costs." The named customers are anonymised as Customer 1, 2 and 3.

Two weeks later they filed an update. It is the most honest revenue line in the library:

"Current revenue = $1968 + ~$100 unbilled usage + (+$399 in transit, payment ran into issues with Stripe)."

Under two thousand dollars booked, with the Stripe failure itemised rather than rounded away.

They had also already lost twice. "We pivoted to the current product last month after two previous YC submissions and two failed ideas." The immediate predecessor was an AI job-search copilot. Their lesson from killing it came in three clauses, one of them just YC's own motto repeated back: B2C recruiting churns, building the wrong thing is expensive, ship in weeks not months.

Andy Li took 51%, Raymond 49%. Andy did design and frontend. Raymond did full-stack with a focus on ML. Andy had worked at Yummy Future (YC S19) in 2021, which is how he had heard of YC at all.

They got in. Winter 2024.

Where they are now

The company is Respan, formerly Keywords AI Inc. In March 2026 it announced a $5 million seed round. Its backers were Gradient, Y Combinator, Hat-Trick Capital, Xiaoxiao Fund, Antigravity Capital and Alpen Capital, alongside angels (SiliconANGLE, 18 March 2026).

The one-line description on its YC page is now "self-driving observability, evals, and gateway for AI agents." Read the order. Observability first, evaluations second, and the gateway (the router the whole application was about) third.

The numbers on that page: more than 100 AI startups and enterprise teams, 1B+ logs and 2T+ tokens processed every month, supporting more than 6.5M end users. Team of 15, San Francisco, YC partner Harj Taggar. The funding announcement names three co-founders — Raymond Huang, Andy Li and Hendrix Liu — so the two-founder company that wrote the application is a three-founder one now.

So the routing did not fail. It got demoted.

The mechanism, because this is the part worth stealing

A router is a proxy. Every request your product makes to a model passes through it. A proxy that decides where a request goes necessarily sees the request: the prompt, the model chosen, the tokens spent, the latency, the response, whether the caller came back. Building the router meant accidentally building the most complete log of an AI application's behaviour that anyone had.

In 2023 the valuable thing about that position was the arbitrage: send the cheap task to the cheap model, undercut GPT-4 by 35%, keep the difference. The arbitrage was real and it was also temporary. The spread between models is set by other people and closes on their schedule. The position itself remained. When agents started failing in ways nobody could explain, the company already had the pipe every agent's traffic ran through.

Their own pitch names the difference. Co-founder Andy Li's argument, given to SiliconANGLE, is that many existing platforms stop at retrospective tracing and scoring rather than helping teams find regressions, root causes and the next step. Respan's evaluation agent, he says, "can analyze failure across different trials, find root causes to specific decisions agents are making, and recommend what eval to add next."

Two pieces of operational advice from the same interview are worth more than most vendor blog posts. Run your evaluating model from a different model family than the one running the agents being evaluated. Otherwise the evaluator's biases are the evaluated system's biases. Set the evaluator to a low temperature, too, so its output stays predictable instead of varying creatively. Both are free to apply tonight, whatever tooling you use.

What the application got wrong, and it does not matter

The competitor answer is confident in a way that reads badly now: "We outperform our competitors in both cost and output quality. As the cost, speed, and quality of individual LLMs like GPT-4 improve, so does our product, ensuring we always maintain a competitive edge."

That last clause is the mistake. Improvement in the underlying models did not reliably improve the router. A good enough cheap model erodes the reason to route at all, and every serious gateway now does the same trick. The market-size line has the same tone: "a huge market size of ~$200 billion for the AI market, our current serviceable available market is $2.7 billion." A number that arrives without a method.

None of it stopped the application from working, which is the useful lesson. YC did not fund the SAM calculation. It funded two people who had killed two ideas in a year and launched a public beta on a Thursday. They booked twenty-plus demos on day one and could tell you the exact dollar a Stripe error was holding up.

Read it knowing that the answer they were proudest of — nobody is combining LLMs yet, and this is where LLM usage is going — turned out to be correct, and turned out not to be the business.

The application

YC Winter 2024 · Accepted · Source

Easily build, deploy, and monitor your AI application. Keywords AI is basically Datadog for LLM applications.

Company

Describe what your company does in 50 characters or less.

API Bartender for LLM APIs.

What is your company going to make? Please describe your product and what it does or will do.

We are building an API that combines different LLMs for quality and cost optimization. Our machine-learning model sorts tasks based on complexity and specialization, directing them to the best-performing LLM in those categories. This approach allows companies to save costs, as cheaper models can perform equally well in certain areas. Currently, we are 35% cheaper than GPT-4 without sacrificing quality. Startups and SMEs use our service to save money and get better results.

Where do you live now, and where would the company be based after YC?

Chicago, US / San Francisco, US

Founders

Please enter the url of a 1 minute unlisted (not private) YouTube video introducing the founder(s). This video is an important part of the application. (Follow the Video Guidelines.)

(no answer)

How long have the founders known one another and how did you meet? Have any of the founders not met in person?

We met through a mutual friend in college and have been best friends since then. We have been living and working together.

Progress

How far along are you?

We launched our public beta yesterday (10/12). We already have paying customers as well as free trial users. We're working on integrating Claude, PaLM, and some fine-tuned models.

---- Update 10/27 ---- Over the past two weeks we have been focusing on four things:

1. adding user requested features, such as organization management, usage tracking, and LangChain integration.
2. addressing user concerns, such as GDPR compliance, terms of use, and privacy policy.
3. sales & user engagement - we booked 23 demos up to today and helped 3 companies with custom integration.
4. QA & testing.

How long have each of you been working on this? How much of that has been full-time? Please explain.

We started our company in January and have been working full-time. We pivoted to the current product last month after two previous YC submissions and two failed ideas.

How many active users or customers do you have? If you have some particularly valuable customers, who are they? If you're building hardware, how many units have you shipped?

We have 17 users running our free trial API key since yesterday's launch (10/12). We also have seven B2B customers who are paying and working on integrating our API. Although we don't yet know precisely how much they will pay, we expect ~$500-1500 /user per month based on their current AI usage costs. The three customers we expect the most revenue from are Customer 1 (B2B knowledge base chatbot, 100+ customers), Customer 2 (SQL copilot, 10k+ users), and Customer 3 (AI for codebase, $10k+ monthly AI spending).

---- Update 10/27 ---- Current revenue = $1968 + ~$100 unbilled usage + (+$399 in transit, payment ran into issues with Stripe). Customer 4 is paying us close to $1800.

Do you have revenue?

Yes

We're interested in your revenue over the last several months. (Not cumulative and not GMV).

All our revenue comes from customer API calls based on usage.

Anything else you would like us to know regarding your revenue or growth rate?

We plan to update our pricing model from subscription-based to solely pay-by-usage, as many users have indicated a preference for this model.

---- Update 10/27 ---- We have updated to usage-based pricing.

If you are applying with the same idea as a previous batch, did anything change? If you applied with a different idea, why did you pivot and what did you learn from the last idea?

We pivoted from an AI job search copilot and learned three key lessons: 1) B2C faces high churn in recruiting. 2) User feedback is vital; building the wrong thing is costly — "make something people want." 3) Quick launches are crucial for finding product-market fit; now we ship in weeks, not months. We believe our current idea is more scalable and can grow very quickly.

Idea

Why did you pick this idea to work on? Do you have domain expertise in this area? How do you know people need what you're making?

We chose this idea after encountering limitations with GPT-3.5 and high costs with GPT-4 in our previous project. The concept actually began as a cost-optimization solution for our own use before we realized its broader application.

With conversations with over a hundred founders in the generative AI space, we have a strong grasp of the domain and customer needs.

We knew people wanted our product when we had more than 20 demos booked on the day we launched. We received many "please consider us for a demo" and "this is such a great product" messages. One customer we talked to today was especially thrilled that we could save them $3,500 monthly on their calls without sacrificing quality.

Who are your competitors, and who might become competitors? Who do you fear most?

Our competitors are companies that offer single or fine-tuned LLMs. We will incorporate these individual LLMs into our product, as well as fine-tune some LLMs for specialized use-cases.

We outperform our competitors in both cost and output quality. As the cost, speed, and quality of individual LLMs like GPT-4 improve, so does our product, ensuring we always maintain a competitive edge. There are no companies working on a combined LLM as of now, and we firmly believe this is the future for LLM usage.

How do or will you make money? How much could you make?

We'll make money with a pay-by-usage API model. Right now our profits are more than half of our revenue. We'll initially target startups and SMBs using generative AI.

With a huge market size of ~$200 billion for the AI market, our current serviceable available market is $2.7 billion.

Equity

Please describe the breakdown of the equity ownership in percentages among the founders, employees and any other stockholders. If there are multiple founders, be sure to give the equity ownership of each founder.

Andy, CEO - 51% Raymond, CTO - 49% We also plan to create a 20% employee pool for early hires.

Legal

Who writes code, or does other technical work on your product? Was any of it done by a non-founder? Please explain.

We both write code. Andy does design and frontend; Raymond does full-stack with a focus on ML.

Others

If you had any other ideas you considered applying with, please list them. One may be something we've been waiting for. Often when we fund people it's to do something they list here and not in the main application.

We currently do not have other ideas; however, as a startup that has pivoted twice, we're adaptive and understand the importance of making decisive changes for higher potential projects.

Curious

What convinced you to apply to Y Combinator? Did someone encourage you to apply?

We firmly believe YC can boost our outcome by at least 7%. We plan to use the 3 months at YC to continue focusing on building our product and talking to users — every generative AI YC startup as a potential customer.

How did you hear about Y Combinator?

Andy worked at Yummy Future (YC S19) in 2021 and learned a lot of good things about YC.