In this edition of Founder Spotlight join us with a conversation with Dominik Rehbock, CEO and Co-Founder of Hyground, for an insightful conversation about his journey as a founder and the thinking behind building an AI SRE agent that can autonomously resolve incidents. The interview explores how he and his team are tackling one of the biggest challenges in modern infrastructure: making operations faster, smarter, and more reliable through automation. The interview highlights the problem-solving mindset, product vision, and innovation that are shaping Hyground’s approach to the future of site reliability and system operations.
This conversation highlights not only Dominik’s vision and leadership but also the broader trends shaping the future of startups and capital management – read on to learn more.
What’s the story behind Hyground?
It started with pain. Our CTO Bastian worked at Bloomberg in operations on high-frequency trading servers. If the server goes down, the costs are enormous. He understood the stress of on-call engineers firsthand. They’re basically emergency doctors for IT. Our CPO Benny spent most of his career in test automation and observability, so he knew exactly where observability fell short.
After a few years working together at an IT consultancy, the idea of Hyground was born. We solve IT incidents within minutes rather than hours, taking the stress off on-call engineers. More than that, we enable reliable, autonomous IT operations. The firefighting gets easier, and the regular work behind the scenes that prevents the fire from breaking out in the first place gets better too.
I joined Bastian and Benny in January 2025 as the third co-founder, and we went full-time at the beginning of that year. My background is a different one. Mechanical engineering by training, then strategy consulting at Bain & Company and a strategy role at INNOLITH. That split turned out to be the right one. They had spent two decades in operations and had a working prototype. What the idea needed next was a company around it: a market to point at, customers who would pay for it, and the money to build. We incorporated Hyground GmbH in Hamburg in June 2025.
The proof then came in the order you want it. Deutsche Bahn ran a proof of concept and became our first licensed customer, and that buy signal is what sent us looking for venture money. We started fundraising in October, signed a term sheet in November, and closed the round in December. We announced the €3 million pre-seed at KubeCon Europe in Amsterdam in March 2026. Partech led the round, alongside Adesso Ventures, Angel Invest and Plug and Play.
The angels on the cap table matter just as much to me. Julius Volz the co-founder of Prometheus, the de facto standard for metrics in cloud-native infrastructure. Volker Maiborn and Holger Wolff are two of the most experienced IT consultants in the German-speaking market. When the people who have lived this problem at that scale put their own money in, you know the problem is real.
Today we are around 16 people, in production at Deutsche Bahn, ifm, automotive manufacturers, and e-commerce and retail companies. Insurance and banking are where most of our new conversations are starting. The round funds one thing: making the agent capable of more of the operational work, one workflow at a time.
What does an AI SRE agent actually look like for an operations team?
Two things. First, on-call incident analysis: the pager fires, and the engineer walks into a solved diagnosis. A structured summary with what happened, the evidence, estimated blast radius, and a proposed fix is already waiting.
Second, the repetitive work that currently pulls senior engineers away from what they should be doing: CVE impact assessments, deployment preparation, rollout planning, cloud cost reporting. Hyground spans an intelligence layer over the IT environment and can access all the same data points engineers can.
That’s why we call it a virtual software operations team, not just an SRE tool. It covers the full breadth of recurring ops work.
You serve regulated industrial operators and high-growth SaaS scale-ups. Why both?
We picked a constraint, not a vertical: environments where downtime is genuinely expensive and throwing more humans at the problem doesn’t scale.
Scale-ups give us velocity. Infrastructure is growing faster than the team, and they can close in weeks. They push us on workflow breadth. Regulated enterprises like Deutsche Bahn and ifm give us depth, longer cycles, and push us on audit quality and legacy coverage.
We can only run both motions because the underlying architecture (self-hosted, inside the customer’s trust zone, auditable by design) serves both segments natively. A generic SaaS AI tool can’t play in the regulated segment. A regulated-only tool can’t move fast enough for a scale-up.
Is AI really the right approach for environments with critical uptime requirements?
It’s not optional. The complexity of IT operations is outgrowing human capability. Gartner named the category this year, autonomous reliability, because human-centric operations have exceeded their capacity to handle modern infrastructure.
In five years, all IT operations will use AI agents the way all developers use AI coding assistants today. We’re seeing that in every customer conversation. The question isn’t whether to adopt. It’s how fast.
How do regulated and industrial requirements shape the product?
They shape almost every architectural decision. Four points, because these are the questions CISOs ask in the first meeting.
Deployment inside the trust zone. Hyground runs wherever the customer’s data already lives: managed LLM inside their hyperscaler, self-hosted models on their own GPUs, or fully air-gapped on-prem. The reasoning layer stays inside whatever boundary the customer has defined.
Read-first by design. The agent investigates and recommends. It does not change your infrastructure on its own. Most adapters are read-only and verify that at startup, so a database adapter refuses to start if the credentials you hand it can write. Where a write exists at all, like a ticket comment or an email, it is narrowly scoped. Any infrastructure write requires explicit human approval per command, bound to that exact command and session, and the system fails closed on anything it cannot classify. We would rather name the handful of places we can write than claim zero.
Full auditability, including the reasoning. Every query, every piece of evidence, every recommendation is logged, including the decision path so a compliance reviewer can reconstruct why a recommendation was produced, not just what.
Scoped, least-privilege access. The agent sees only what the customer grants it, per cluster, per namespace, per data source.
How do you ensure operators can trust an AI recommendation?
Every diagnosis is backed by the specific data the agent used: the exact log line, the exact metric spike, the exact commit. The operator can click through and verify without redoing the work.
The entire working process is visible. Operators see the agent’s to-do list in real time and the exact commands it ran. Nothing is hidden.
What surprised us: this turned into a teaching tool. Junior engineers watch the agent work and learn how a senior SRE approaches the same problem. In organisations struggling to hire and retain senior talent, that knowledge-transfer effect has become a major reason to deploy us.
What’s the most valuable output for a customer?
Faster triage and better root cause analysis are where the immediate value lands. Operators walk into a pre-built diagnosis instead of a blank screen, and time-to-understanding moves visibly in the first weeks.
We deliberately don’t quote aggregate percentages. Every environment produces a different number. We scope every PoC with customer-agreed success metrics before it starts: time-to-diagnosis, operator hours returned per week, share of cases where the agent matched the senior engineer’s recommendation.
The bigger story is what happens besides on-call incident resolution: the endless, repetitive work that keeps IT operations teams busy. Incident response is the one workflow we have taken end to end so far, and the roadmap is to take the rest off engineers one at a time. We are deliberately strict about what counts as shipped. A workflow only counts when it is a reusable procedure, proven in a real customer environment, and wired to production triggers. The strictness is the point. It is why the ones that ship actually hold.
What objections do enterprise buyers raise most often?
Three come up in almost every evaluation.
Data and compliance. It’s not that enterprises can’t send telemetry to a third-party SaaS. It’s that they don’t want to connect the heart of their IT system, meaning the virtual machines and the Kubernetes clusters, to a SaaS player that would have access to everything they have. For most AI ops tools, that’s a hard blocker. For us, it’s where the conversation gets easier, because the entire stack, including the reasoning layer, runs inside the customer’s trust zone.
Control. “What can the agent do, and can we stop it?” Read-only today, scoped access, and the customer holds the kill switch. That takes the scariest part of the conversation off the table on day one.
Integration. “Can it connect to our stack?” Yes. Hyground integrates with AlertManager, observability platforms, Confluence, Git, ticketing systems, and event pipelines out of the box. If a connector doesn’t exist yet, we build it.
Your advice to enterprises still cautious about AI in operations?
Start small, start observable, and pick a use case with an obvious measurable outcome.
Choose one recurring, painful workflow (a CVE assessment, a monthly cost report, a specific class of recurring incident) and deploy the agent in read-only, advisory mode. Agree before it starts what success looks like. Let it run alongside for a few weeks. Measure.
The mistake is treating autonomous operations as a big-bang transformation. It’s progressive delegation, earned one workflow at a time, with the operator keeping control at every stage. The companies that get this right in the next two years will have a real cost and resilience advantage over the ones that don’t.
