ML Research Scientist - Video AI

Overview

We're building AI systems that generate video ad creatives at scale. The hard problems here are genuinely hard - realistic avatar generation that avoids uncanny valley artifacts, natural-sounding audio synthesis, seamless product integration without visual glitches or temporal inconsistencies. This is the space you'll be working in.

What Matters To Us

Background

CS degree from a strong program. 10+ years of experience. You've spent meaningful time at companies known for serious engineering and research - places where the bar is high and the problems are real.

Video Generation & Fine-tuning

You've worked hands-on with state of the art video models. Not API wrappers - actual fine-tuning, architecture decisions, training pipeline design. You understand techniques like IP-Adapters and LoRAs and have applied them to real problems around avatar realism, audio quality, and visual consistency in generated video content.

Agentic Video Editing Systems

A core part of this role is designing and building autonomous agents for video review and editing. Think systems that can identify artifacts, evaluate quality, and make intelligent editing decisions. We'd love someone who would naturally think about defining benchmarks for this - how do you evaluate a video editing agent? - and then publish that work.

Evaluation in Non-verifiable Domains

Video output can't be unit tested. You've thought deeply about evaluation methodology when ground truth doesn't exist - perceptual metrics, human eval pipelines, automated quality scoring, A/B testing frameworks. You have a real point of view on this.

End-to-End Ownership

Research to production. You own the MLOps - training infrastructure, deployment, monitoring, iteration. No handoffs.

Performance Feedback Loops

Ad performance signals - CTR, CVR, engagement - should inform what gets generated. You'll architect systems where downstream metrics feed back into the generation pipeline.

AI-native Engineering

You push AI coding tools to their limits. But you're a strong enough engineer to know when the output isn't right and write production-grade code yourself.

Publishing & Thought Leadership

You write and publish - white papers, technical articles, benchmark results, open source contributions. You can articulate what you're building, why it matters, and what you've learned. This is important to us.

---

We need depth. If the extent of your generative AI experience is prompt engineering, this isn't the right fit.