Shubham Palriwala spent years listening to open-source communities and analytics dashboards. Now his company, Agnost AI, listens to a million machine conversations a day - and quietly ships the fixes.
There is a moment every founder remembers - the one where the product stops working and nobody knows why. For Shubham Palriwala and his co-founder Parth Ajmera, that moment did not arrive as a bug report or an angry email. It arrived as silence. An AI product they had built was slipping away from them, and to figure out what went wrong they did the least glamorous thing imaginable: they opened their Discord and started scrolling.
Somewhere in that log was the whole answer. Users had been telling them exactly what they wanted, where they got stuck, and why they left. The company already owned the transcript. It just was not reading it. That realization - the insight was already there, unsurfaced - is the seed of Agnost AI.
With AI agents, there's no complaint - just churn.
It is a tidy line, and like most tidy lines it hides years of pattern recognition. A traditional app annoys you and you file a ticket. An AI agent disappoints you and you simply stop coming back. No signal, no warning - just a number that drifts down. Palriwala's bet is that the signal exists anyway, buried inside the conversations themselves, and that a machine can be taught to pull it out and act on it.
Long before the pitch deck, Palriwala was the kind of engineer who collected open-source programs the way other students collect internships. He grew up in the world of Summer of Bitcoin, Google Summer of Code, Google Season of Docs, and Linux Foundation X - contributing to OWASP's Juice Shop, OpenMined's privacy tooling, and Kubernetes-adjacent projects like Kyverno. He earned a Certified Ethical Hacker credential and, in a small hacker in-joke, wrote his GitHub bio as a Unix fork bomb.
He studied computer science with a specialization in information security at the Vellore Institute of Technology, graduating in 2023. But the resume line he tends to lead with is a different kind of credential: he was the youngest engineer on Cisco's analytics team, where he worked on reactive programming and performance engineering with the Cisco Spaces group and co-authored a provisional patent on eBPF and WPA3.
Then came Formbricks, the open-source answer to survey tools like Typeform, where Palriwala was Engineer #1 - the first person hired into a codebase that would grow to more than twelve thousand GitHub stars. Being the first engineer at an open-source company is an unusual education. Your users are visible, vocal, and everywhere: in issues, in pull requests, in Discord threads. The whole job is listening at scale. It is not hard to draw a straight line from that to what he builds now.
Building an MCP server isn't just wrapping endpoints - it's about designing for how models actually think and work.
Palriwala and Parth Ajmera are not new acquaintances who met at a hackathon. They have been friends for roughly eight years, going back to childhood, which means the co-founder trust that most startups spend a year building was already there on day one. Ajmera brings a complementary sharpness: he ranked 159 out of 1.5 million candidates in IIT Madras's computer science program and went on to build terabyte-scale data pipelines at Microsoft.
When their earlier AI product stalled, the pair went through Entrepreneurs First - the talent-first startup builder - as part of its S25 cohort, raising $250,000. But the more important thing they walked away with was the observation from that Discord scroll. They had been sitting on a goldmine of user intent and treating it like noise. So they built the thing that would have saved the last company.
Agnost AI describes itself as a self-improving layer for any AI agent in production. In practice it wraps an agent and runs three loops at once. It is a genuinely different shape from a dashboard, because the point is not to show you a chart - it is to close the loop without a human rewriting prompts every cycle.
Every input and output is parsed for what the user actually wanted and how the agent responded.
Production traffic is scored in real time against guardrails and quality criteria.
Insights feed back in - opening pull requests against prompts, tools, and harnesses.
The system clusters recurring intents, feature requests, frustration, and failure patterns, then can spin up an evaluation test or open a debugged pull request against the agent's prompts and tools. It processes more than a million AI events a day and connects in three lines of code. In the summer of 2026 the company joined Y Combinator's Summer batch, with Tyler Bosmeny as its partner, running engineering across San Francisco and Bengaluru.
If you squint, Palriwala's whole career has been one idea rehearsed in different keys. Contributing to open source is listening to a community and shipping the improvement. Analytics at Cisco is listening to a system and finding the bottleneck. Being Engineer #1 at Formbricks is listening to users at volume. Agnost AI is all three at once, pointed at AI agents - which, conveniently, never stop talking.
The company already had the transcript. It just wasn't surfacing the insights.
Palriwala still writes and speaks about the unglamorous plumbing of this new era - how to improve a Model Context Protocol server, how to get more usage out of one, what metrics actually predict whether an agent will stick. It is the same instinct that made him a good open-source contributor: the belief that most of what you need to know is already written down somewhere, if only someone bothered to read it carefully.
His aspiration, stated plainly, is for every AI agent in production to become self-improving - to turn the silent churn into open pull requests. Whether Agnost AI becomes the default layer for that is a story still being written. But the premise is hard to argue with. Your agent has been telling you what's broken the entire time. The only question is who's listening.