Co-Founder of Prized | Software Developer | C++ Developer | Python Programmer | Teaching Assistant | Coach
Every company that tightens its rules to stop shadow AI tools ends up creating more of them. Add an approval step, and someone quietly builds a workaround. Add a review gate, and a spreadsheet nobody vetted appears somewhere in the org chart’s blind spots. Marinos Eliades has built a company on the wager that this pattern is not a coincidence. The rules people write to feel safe are often the reason the unsafe thing keeps happening.
A Mechanical Engineer in the Security Stack
Marinos Eliades is the co-founder of Prized, a Y Combinator S26 company building secure AI infrastructure for internal tools, based in the San Francisco Bay Area. He is a mechanical engineer by training who now spends his days solving a problem usually left to security architects: letting anyone inside a company describe a tool in plain language and have AI build it, without opening a door nobody can close. His conviction is unusual for someone this early in his career. The rules themselves are often the risk.
Trained to Draw the Line First
None of this started with software. In Cyprus, Eliades won gold at the British Physics Olympiad and took first place at the ExxonMobil Sci-Tech European Challenge, the kind of results that come from someone who treats a hard constraint as an invitation rather than an obstacle. Between finishing school and starting university, he served as a Lance Corporal in Cyprus’ National Guard, a stretch of mandatory duty that left him comfortable operating inside rigid structures long before he ever had to design one of his own.
At Stanford, he chose mechanical engineering, working across dynamics, fluid mechanics, and thermodynamics. The discipline runs on one unbreakable habit: define a system’s boundaries and failure modes before you ever try to make it faster. You do not push an engine past its limits until you know exactly where those limits sit. He was later selected as one of thirty students out of three hundred applicants for the Stanford ASES Bootcamp, a twelve-week program that ended with him pitching a design concept to a panel of Silicon Valley investors, an early rehearsal for the fundraising he would do for real a few years later.
That same instinct followed him to the Hong Kong Productivity Council, where he spent a summer on the Green Transportation team, working on projects that ranged from electrifying tourist boats to designing self-sustaining hydrogen fueling stations. It followed him back to Stanford too, where he spent nearly two years as a teaching assistant for CS106A and CS106B, the university’s two largest computer science courses, running Python and C++ workshops for a combined enrollment of more than 1,200 students.
“Mechanical engineering taught me to define boundaries and failure modes before optimizing performance, while teaching computer science reinforced that systems must remain understandable to the people using them,” he says. By the time he took a three-month engineering role at Gander, a Y Combinator-backed aviation AI company in New York, the pattern was set. Define the failure modes first. Build for speed second.
Building Prized Inside a Governed Boundary
Marinos Eliades founded Prized as part of Y Combinator’s S26 batch, building a platform that lets non-engineers build, deploy, and share internal tools with AI, without needing to know where company data lives or how to connect to it. Teams describe what they need. Prized builds on top of the right data, inside what he calls a governed boundary. Without that boundary, companies accumulate what the industry has started calling permission debt: shadow tools, inconsistent access, and brittle integrations that nobody quite owns.
The problem is a specific and uncomfortable one. A useful internal AI tool needs private data, untrusted generated code, and communication with outside systems, all at the same time. The obvious fix is to hand the tool broad credentials and trust a prompt, or an AI judge, to keep it honest.
“The obvious approach, giving the app a database URL or OAuth token and relying on prompts or an LLM judge, fails because a request can look valid while still leaking data,” he says.
The equally obvious overcorrection, a human approval for every action, fails for a different reason. It rebuilds the engineering backlog that pushed teams toward AI tools in the first place, and it drives people straight back toward the shadow tools it was meant to stop.
“Most teams mistake governance for adding approvals, but blanket gates create the shadow tools they are meant to prevent,” he says. “The governed path must be easier than bypassing it.”
Prized’s answer is to make the tool and its security boundary the same object, rather than two separate concerns bolted together after the fact.
“We made each tool and its security boundary the same object,” he explains. “Every tool receives its own Postgres schema and role, so database grants, rather than the model’s interpretation, determine exactly what it can access.”
Generated code never touches a raw credential. Requests move through authenticated gateways and a controlled proxy that only reaches allow-listed destinations. Integrations start read-only. Anything destructive requires a human to approve it first, and none of it depends on a builder remembering the right practice, because none of it is optional.
“Our operating principle is guardrails, not gates,” Eliades says. “Let people move freely inside an explicit capability boundary, then require review only when they expand that boundary or attempt something irreversible.”
Because those guardrails live in the infrastructure itself, a builder inside Prized can iterate without waiting on a security review every time. The company treats every sandbox as disposable and potentially compromised from the start, then keeps the damage contained through defaults rather than vigilance, so that whatever happens inside one tool, the blast radius stays small and predictable.
The Undefined Limit Was the Danger
The instinct to add more rules when something feels risky is almost universal, and almost always wrong in the way Marinos Eliades means it. More approvals do not make a system safer if they make the safe path slower than the unsafe one. What makes his approach different is not that it trusts people less. It is that it removes the need to trust anyone’s judgment in the moment at all, replacing it with defaults that hold whether or not anyone is watching.
He did not arrive at this idea from a security background. He arrived at it from an engine that could explode if you pushed it past a limit nobody had bothered to define. The rules were never the danger. The undefined limits were.
Marinos Eliades is the co-founder of Prized, a Y Combinator S26 company based in the San Francisco Bay Area. He builds secure AI infrastructure that lets non-engineers create and deploy internal tools inside a governed boundary, without compromising security or control. To connect with Marinos or learn more, visit prized.dev or find him on LinkedIn at his LinkedIn Profile.
