Maor Bril

Maor Bril is a software engineer working at the intersection of AI-generated video, infrastructure, and autonomous coding tools. His work at Character.ai tackled a central problem for creative AI: determining whether generated videos tell coherent stories, maintain consistent characters, obey physics, and synchronize sound with action.

Earlier in his career, Bril worked on infrastructure at Coinbase, Netflix, Google, and VMware. At Character.ai, his work has included data pipelines and products such as Stories, Streams, and an AI social feed.

Making video evaluation practical

Bril treats generated-video evaluation as a storytelling problem. Conventional image metrics can assess individual frames, but cannot reliably determine whether characters remain consistent between shots, movement looks physically plausible, or scenes unfold at a believable pace. His approach combines several complementary techniques:

• JudgeJudy and human calibration: His multimodal evaluation harness combines visual metrics, AI judges, and human annotations across specific dimensions, continually adjusting automated judgments against human preferences.
• Pairwise preference comparisons: Asking reviewers which of two videos tells a better story produces more dependable training signals than assigning subjective numerical scores to videos individually.
• Evaluation inside the generation loop: Checking initial frames and short clips for character drift or other defects lets agents repair problems before they spread across an entire sequence.
• A compact Qwen-based video judge: Bril distilled a larger evaluation process into a smaller vision-language model, accepting a modest quality tradeoff to reduce latency and serving costs.

An early version rewarded glossy footage while overlooking static camerawork, implausible physics, and weak storytelling. Bril addressed this by pairing real and generated footage while matching encoding and annotation procedures, helping prevent the system from becoming a synthetic-video detector instead of a quality judge. His AI Engineer presentation on generated-video evaluation also addresses audio-event synchronization and the unresolved difficulty of lip synchronization for animated characters.

Coordinating coding agents

Bril created Clauder, an open-source coordination layer that gives coding agents shared discovery, messaging, and persistent memory across development environments. He has described using coordinated agents alongside independent automated code review.

His companion project agentic routes coding tasks across model providers, tracks spending, applies budget limits, and supports blinded comparisons of generated code. Together, these projects apply the same priorities that shape his video work: coordinated autonomous systems, measurable output quality, and practical control over inference costs.