In a new episode of the PMF Collector, Thibault Jaigu, co-founder of Requesty, sits down with our Partner Matthieu Vaxelaire to talk about the company he almost didn’t build.
Jaigu set out to build a product-analytics layer for LLMs. Five years at Amplitude had taught him that you can’t improve what you don’t measure.
What he actually built, almost by accident, was an AI gateway: the routing and auth layer he and his co-founder needed just to ship their real product, and one they’d sunk hundreds of unpaid hours into before realizing anyone else needed it too.
What follows is a story about the discipline to notice when your own workaround has quietly become worth more than your pitch.
🎧 Full episode below.
1) The problem hiding behind a feature request
Thibault Jaigu spent five years at Amplitude, helping digital-native companies measure retention and engagement. It was analytical work, the kind that trains you to ask “how do we know this is actually working?” about everything you ship.
When GPT-3 landed, he asked the same question about language models. Companies were shipping AI features with no real way to tell if the answers were any good. His idea: a product-analytics layer for LLMs, labeling every prompt and response so teams could track coherence and quality over time.
“Why isn’t there a product analytics tool for LLMs over time?”
2) Fifteen conversations before a single dollar
Jaigu and a close friend spent weeks testing the idea before either of them quit their jobs. They talked to roughly fifteen companies that had already shipped some form of AI, mostly in customer support, and asked how they were judging output quality.
The answer, over and over, was: badly.
“They really didn’t have good ways to evaluate how the AI performed.”
That was enough to validate the problem. It wasn’t enough to get paid. When they finally quit, in September 2023, they had zero paying customers and zero revenue.
3) The whiteboard that told them what business they were actually in
The two of them raised a modest £400,000 pre-seed from angels they already knew (Amplitude’s former first engineer, a couple of ex-COOs) enough runway to build and test without panicking. They brought in a third co-founder, Daniel, whose enterprise background at Palo Alto Networks would matter later.
To actually deliver the analytics product, they had to build real infrastructure: a labeling pipeline, GPUs, data residency in both the EU and the US. Somewhere inside that build-out, without really planning to, they built their own AI gateway — pure internal plumbing to make the real product work.
Months later, they sat down and counted the hours. Literally: tickets logged, hours tallied, on a whiteboard.
“We’d spent hundreds of hours on this ourselves. That’s a real problem.”
The analytics layer they were trying to sell wasn’t the real problem. The gateway they’d built to support it was.
4) Finding out everyone else had the same problem
Around the same time, tool-calling got good enough that demand for their evals product started to evaporate - once an agent could take a concrete action (open a ticket, issue a refund), teams cared less about scoring the raw text of a response. Jaigu and Daniel started talking to more startups to figure out what they actually needed, and heard the same complaints on repeat: rate limits, models unavailable in the wrong region, config chaos wiring up a dozen different provider APIs.
Requesty had already solved all of that internally.
They found their first real traction inside the open-source dev-tool ecosystem '(Cline, Roo, the AI SDK, Mastra) plugging in as a default model provider alongside OpenRouter and the model vendors themselves.
“It’s much nicer to have one API key for two hundred models.”
5) Nice to have vs. must have
The distinction that reshaped the company was almost embarrassingly simple. The evals product was useful, but skippable - a nice to have. A gateway wasn’t.
“If you want to connect your app, you always need a gateway.”
Any team wiring an application to an AI provider needs some kind of routing and auth layer in between. Requesty had already built one, for itself, without setting out to.
They gave themselves one month to turn that internal tool into a public product.
6) The pivot that cost them a co-founder
The one-month deadline came at a cost. One of the three original co-founders, not drawn to the new, dev-facing direction, left for a strong offer elsewhere - while the company was still tiny, with no traction yet on the new idea.
Jaigu and Daniel rebuilt with two. The new version shipped publicly in December 2024. The first month brought in sixteen to twenty accounts, mostly individuals and small teams, with no real distinction yet between a solo developer and an enterprise account. By January, growth had turned a corner.
The business model was simple by design: take roughly 5% of a customer’s AI API spend routed through Requesty, and separately mark up the inference Requesty resells wholesale to most of its clients - margin on both sides of the transaction.
7) Two founders, two playbooks, one differentiator
Jaigu had spent years in enterprise sales; Daniel came from an enterprise background at Palo Alto Networks. That combination shaped the go-to-market early: win individual developers first through the open-source ecosystem, but build toward enterprise governance from day one - letting a company control where its models run, how they’re used, and under what budget.
“We had a lot of enterprise knowledge, and that became our differentiator.”
It paid off fastest in Europe. Many of Requesty’s earliest individual users turned out to be developers at large companies, who wanted their own side projects hosted with zero data retention in the EU - even on a personal account. That governance layer, built for individual peace of mind, became the same thing enterprises would later pay for at scale.
8) The deal that almost didn’t happen
Requesty’s biggest enterprise win followed exactly the land-and-expand pattern their dev-first strategy was built for: one engineer at ZoomInfo, a 1,300-person engineering org, started using Requesty on the side, liked it, and told his boss. That turned into a formal evaluation against serious, well-funded competition - Cloudflare and Databricks.
Jaigu and Daniel spent the three-week evaluation window shipping seventeen new features the client had asked for, barely sleeping.
They won it. Around the same point, Requesty had crossed roughly seven figures in annual revenue from developers and startups alone, and closed a $3 million seed round led by 20VC and Tapestry VC.
The takeaway
Requesty’s story isn’t really about spotting a market before anyone else. Plenty of people saw the AI-gateway opportunity - OpenRouter got there first, grew faster, and was acquired by Stripe for more than $7 billion in August 2026, a constant, slightly uncomfortable point of comparison for Jaigu’s team the whole way through.
What Requesty had, harder to copy, was the discipline to notice when their own workaround was worth more than their pitch. They didn’t find the gateway idea in a market report. They found it in their own ticket backlog, on a whiteboard, counting hours they’d already spent solving a problem they hadn’t thought to sell.
Most founders build the thing they set out to build. The more interesting version of this story is a team willing to throw that thing away, once the data made it obvious the byproduct was the business.
