Empirik announced on Tuesday that it has spun out of Sequoia Capital as an independent company with $21 million in seed funding. Its origin is unusual: it started inside the venture firm's own IT organization. Avon Puri, who joined Sequoia in 2020 after more than a decade working on infrastructure at Rubrik and VMware, began it three years ago with Sudheer Dhurjati, another IT leader at the firm, when large language models started to show what they could do. The product watches changes across large infrastructures and works out what each change could cascade into. The pitch is prediction rather than response: catch the failure before the systems go down instead of reacting after they have.
Sequoia treated it as a new category of observability tool — one that prevents incidents rather than explaining them afterwards — and developed the project through 2023. The firm then brought in Kartik Chandrayana, who was appointed chief executive earlier this year.
Bogomil Balkansky, the Sequoia partner, frames the gap this way: companies have spent heavily for years on keeping systems up, but most observability tools do not understand the complex dependencies between components. Empirik, he says, is one of the first purpose-built platforms for tracking change across infrastructure at scale. It behaves as an autonomous dispatcher — low-risk changes pass, larger ones get constraints attached, and the most dangerous updates are escalated to a human. The intended effect is that DevOps teams and site reliability engineers hand off part of the routine hunting and fixing and spend their time elsewhere.
Customers since the launch earlier this year run from young companies to several members of the Fortune 500, including S&P Global, Guardant Health and a large consumer goods company. Chandrayana argues that as software development accelerates, AI tools that help infrastructure engineers cope with constant change become steadily more necessary, and that Empirik wants to do for infrastructure engineers what AI systems have done for software developers. The company's stated goal is to automate part of an infrastructure engineer's work the way Cursor and Claude Code automate individual developer tasks.
That analogy is the weakest part of the story — my judgement, not the company's. A coding assistant works where being wrong is cheap. A bad suggestion costs a keystroke; the engineer is looking at it, the loop is a second long, and the blast radius is a file. An infrastructure dispatcher operates where both errors are expensive in opposite directions. Wave through the change that takes production down and the tool has done worse than nothing. Block or constrain a safe change and the tool becomes the reason a release is late. Cursor's economics do not transfer to that position, and the dispatcher design tacitly concedes it by routing the riskiest changes back to a person.
The other thing to hold in view is that every load-bearing claim here comes from the same address. Sequoia employed the founders, incubated the product, seeded the round, appointed the chief executive, and supplies the partner who defines the category, diagnoses the market and draws the competitive map. Balkansky places Empirik in a category of its own, complementary to AI platforms for site reliability engineers such as Resolve and Sequoia-backed Traversal — a boundary drawn by an investor on both sides of it. None of this is improper; incubation works exactly this way. It does mean the assessment of how well Empirik works is currently an internal one.
Which leads to the question the announcement leaves alone: how anyone tells whether the predictions were right. A forecast failure that gets prevented looks identical to a false alarm, and there is no number anywhere in the launch — not a detection rate, not a false-positive rate, not an incident count before and after — attached to S&P Global or anyone else. The whole category has this problem, and a product built entirely on prediction has it most acutely.
So what Empirik is really selling its customers is the sorting, and the sorting is judged by events that never happen. That works until it does not. The first time a change the dispatcher waved through takes a Fortune 500 system down, the accounting will be exact, and it will run in only one direction.