After quitting our jobs and failing twice, we finally built a product doing $5k MRR and got into Y Combinator. But halfway through the batch, we pivoted. It was summer 2024. Sidhant, Shashij, and I were launching ideas in parallel. At one point, we had 4 running. One was Lex, which helped insurance lawyers find medical records. It reached $5k MRR and got us into YC. Voice AI testing was another. The more time we spent on it, the more convinced we became that it was a much bigger problem, and one we were far more excited to spend years working on. So we changed direction during the batch. I think founders can overvalue traction. Revenue proves an idea is working. It doesn’t prove it’s the best company for you to build. Working on multiple ideas gave us enough distance to see that. It also made it easier not to force something just because it was the only idea we had. Paul Graham says to treat a startup idea as “a hypothesis rather than a blueprint.” That’s how we found Cekura. We kept putting ideas into the world and stayed willing to change our minds, even after one started making money.

August was @cekuraAi 's biggest month yet. Voice agents are getting good. We ran 9 production agents through the same 10 scenarios, 3 times each, 270 calls in total, and almost nobody failed on reliability. Two years ago reliability was the only thing anyone measured. Where the agents separated was humanness, tone, latency, and what happens when a human talks over them. @usebland was the fastest to shut up when interrupted. @retellai had the lowest turn latency. I think a caller notices those things long before they notice whether the task technically got completed. It's also why enterprise deals are closing faster for us. It matters less now whether voice AI works because everyone knows that it does, what matters now is whether it'll still work after a prompt change or a provider update. Two product things I'm proud of. Cekura now connects to GitHub and opens PRs with fixes when a test fails. Finding the bug was never the hard part for our customers. Getting the fix into the repo was. You can also replay a real production call with the original audio, background noise, and timing. Trying to reproduce a voice bug from a transcript is mostly guesswork. Replaying the actual call removes most of it. We also partnered with Speechmatics, Reson8, and Inworld, went to Genesys Xperience and Dreamforce, and shipped a pile of smaller things (monitor presets, redaction, auto-refunds when a provider drops the call). Speech-to-speech is going to make all of this harder. OpenAI and Google both shipped new models, and my read is that these systems give you fewer places to inspect what happened when something goes wrong. Which is good for us. Thanks to the team and our customers. We have a lot in the works for September. Benchmark: benchmarks.cekura.ai/experim…

Build something people want they said. Make a lot of noise they said @cekuraAi is now the infrastructure for self-improving agents. It catches the failure, hands it to your coding agent, and ships the fix as a PR. We're live on @ProductHunt today. An upvote helps a lot, and I'll take any feedback