Principal AI Engineer

About Sondera

AI agents act on what people write, not on everything they mean. Each step can be allowed while the sequence does harm. Sondera's policy engine decides each action from everything the agent has done in the session, using natural-language rules we turn into formal policy. The same action, in the same state, under the same policy, gets the same decision, whichever vendor built the agent. The open problem underneath is semantic integrity: every action carries out what an authorized person meant. We've published our work on it at an ICML workshop and presented it at Black Hat Arsenal, BSides and [un]prompted. Our open-source hooks and policy engine already govern coding agents, and enterprise customers run our platform. Our team has built products at Cylance, Databricks, HiddenLayer, NetRise and Obsidian Security, and we're backed by Decibel, Runtime Ventures and Aviso Ventures.

About the role

You turn what AI agents do into the policy that governs them. Our policy engine decides each action by rule. Your models read agents' trajectories across whole fleets, see what rules can't, and propose the policies and repairs a security engineer approves. You decide how monitors are evaluated, where a model's judgment stops and a rule takes over, and which proposals are worth a person's review. You'll turn our AI security research into enforceable, systems-level controls.

Principal is a senior individual-contributor role; you lead through design and review.

In your first year you will

• Build labeled evaluations from real agent activity, and publish each monitor's precision and recall.
• Run monitors across every customer's fleet at a cost per trajectory we can predict, and show which events they haven't read.
• Turn patterns across many trajectories into proposed policies, each citing its evidence and replayed against recent activity.
• Turn denials, approvals and overrides into proposed refinements, and simulation failures into proposed repairs.
• Map our policies to control frameworks, with the evidence for each mapping.
• Publish or present what you learn at an ML or security venue.

You may be a good fit if you

• Have shipped LLM-based judges, classifiers, monitors or agents in production, with evaluations you trusted.
• Build evaluation sets with human labels, and report results with their uncertainty.
• Write production code in Python and a systems language.
• Design systems where a model proposes and something else decides.
• Have a public record of research: papers, talks, CVEs or open-source work.
• Use coding agents daily, and review what they write as closely as a colleague's work.

Prior security experience is useful but not required. Knowing when not to trust a model's output matters more.

Strong candidates may also have

• Research on AI control, trusted monitoring or scalable oversight.
• LLM judges for behavior that spans many steps or many agents.
• Systems that learn from reviewers' decisions.
• Mapping controls to frameworks.
• Red-teaming of LLM systems, including prompt injection.
• Production Rust experience.

Our stack

Rust for the services that render verdicts, Python for AI and research, and TypeScript for the web applications. Policy compiles to a formal policy language. The platform runs in Google Cloud and in customers' own infrastructure.

How we work

The people who write our papers ship the product, so what breaks in production sets our next question. We study four open problems: turning written rules into policy, enforcing that policy at runtime, making agent behavior inspectable, and overseeing fleets of agents without a person reviewing every action. We build Sondera with coding agents it governs, and a person who owns the code approves every change they make to production. Whoever builds something runs it. We work remotely and in writing, and we flag decisions we can't undo before we make them.

Logistics

• Location: Remote, United States. We meet in person three times a year.
• Compensation: Base salary of $220,000–$260,000 USD, plus equity.
• Benefits: Medical, dental and vision coverage for you and your dependents; a 401(k) plan; unlimited paid time off, and we expect you to take about a week each quarter, plus paid holidays.
• Visa sponsorship: We sponsor visas, including transfers, though not for every role or candidate.
• Interview process: screen → work-sample interview and review → research discussion → career deep dive (a structured walk through your past roles) → references. We aim to make a decision within two weeks of your screen.
• Using AI in your application and interviews: Use AI to prepare, and to polish an application you wrote yourself. In the work sample you'll build without AI, then extend your work with a coding agent, so we see both how you design on your own and how you check an agent's output. The other interviews are without AI.

We encourage you to apply even if you don't meet every qualification.