ML Engineer, Generative Media

VideoClaw is a San Francisco applied lab for AI and creativity. Its first product brings prompt-to-video to a Mac app: thousands of AI models, web search, motion graphics, and video editing inside Claude or ChatGPT, using the person's own GPU so the work stays private, cheaper, and smarter. The launch video drew 1.7 million views. The idea is that video is the world's most widely understood language, and few people can speak it. The founding team has built and sold two companies in consumer and media, and one of them built an AI product to about $500,000 in annual revenue alone. Backers include marquee names, named only at the end of the process. Seed stage, about 1 to 10 people, founded in 2026, in San Francisco.

VideoClaw is hiring an ML engineer to own the model side of its video harness. The company is starting to train its own small models so the harness gets smarter and faster, including proprietary work. You own training and evaluation for the models used inside the product. You build benchmarks and feedback loops across video-generation models. You turn product usage and human creative judgment into high-signal training and evaluation data. The instinct that matters is knowing what data to collect, label, filter, and feed back so model behavior actually improves.

This hire weights the company and the specific work far more than school. They want industry implementers who build and ship. A deep research background can be considered, and the person still has to ship, not only advise. Any strong modern generative-media background works: diffusion, flow matching, and video or loop models. The method can be supervised learning, reinforcement learning, or something else, as long as it is generative media.

You own training and evaluation for the models inside the harness. You own the benchmarks and feedback loops. You turn usage and creative judgment into training data. You own the proprietary small-model work that makes the harness smarter and faster.

Bar: about 3 to 6 years. You have trained, fine-tuned, and shipped useful models end to end. You have hands-on depth in diffusion, flow matching, or other modern image and video generation. You have a real feel for training-data strategy. The backgrounds they name are Meta, ByteDance video models, Google video models such as Veo, or a comparable frontier video or generative-media lab. A named video model you trained or shipped, a frontier-lab background on video, and benchmarks or training-data pipelines are green flags. A pure researcher who does not ship, classical ML with no generative-media depth, no recognizable lab on the work, or no feel for training data is a miss.

In office in San Francisco. Visa sponsorship is available. Equity is about 0.1% to 0.7%. Posted work intensity is 9-9-5. The process is an intro with the CTO, an on-site work trial, a culture conversation with the CEO, then references.