Multimodal models that read video moments, adversarial ad creative, and interactive experiences.
Content Understanding · AI Moderation · Agent Governance
I build AI that understands content — and the systems that govern what it does.
I'm Ziyad Mir. I build content understanding and moderation AI at Roblox. The models read ad creative and interactive experiences, and what they learn feeds signals across the platform — moderation, brand suitability, discovery, ranking, creator and user surfaces. My job is making those models trustworthy enough to act on: evals, policy decisioning, human escalation, and audit trails behind every call.
Before Roblox I was at YouTube, where I built moment-level video understanding for brand advertising. YouTube launched it publicly as Peak Points in 2025 (CNBC, TechCrunch), and the underlying signals have since spread through the platform — they now power contextual sponsorship matching between brands and creators. The problem I'm taking on now: understanding what actually happens inside Roblox experiences, by training agents to go play them.
Moderation and policy decisioning: evals, permissions, human escalation, auditability, enforcement.
Runtime substrates and deployment control for AI that acts: sandboxes, tool access, rollback paths.
Production evidence across ads, discovery, moderation, and high-throughput platform infrastructure.
What I'm working on now
My current focus is the layer where AI meets real users on a platform with a young audience: the models, decisioning, evidence, and operations a system needs before its judgments can be trusted.
- Moderation systems: pipelines that convert creative, audio, destination, and policy evidence into auditable, enforceable decisions.
- Content understanding for paid and organic surfaces: brand suitability, age appropriateness, discovery quality, and measurement.
- Runtime substrates for agents: sandboxed execution, tool access, artifact streams, task state, and rollback hooks.
- The research questions underneath all three: how to evaluate models whose judgment calls carry policy and safety consequences, and how agents can read content from the inside.
How I think about the problem
Three claims run through everything I build. They are also the research agenda: each one is easy to state and hard to make true in production.
Understanding decides what a system can see
Video moments, adversarial creative, interactive worlds — every platform decision is bounded by how well the models read the content underneath it.
Governance decides what it is allowed to do
Policy, safety, and business risk become system contracts: eval gates, permission matrices, human escalation, audit trails, and rollback paths.
Autonomy has to earn scope
Start with controlled validation, measure disagreement and incident rates, then widen what the AI may decide only where evidence supports it.
Explore the work
Different ways to understand my approach and execution style.
Projects
Case studies across content understanding, AI moderation, advertising systems, and platform infrastructure — including the YouTube Flash → Peak Points arc.
Writing
Technical essays on deployment control, agent runtimes, evaluation, content understanding, and how AI changes engineering work.
Notes
Research notes on open problems: experience understanding via in-sandbox agents, evaluating judgment-call models, deployment control.
Life
A running memoir — Saudi Arabia to small-town Ontario to Waterloo, and the career that followed: Quora, Opendoor, Uber, Google, Roblox.