Velvet

Skills: Operations, Python, SQL, Claude Code Velvet is a data research company founded by Lucas Mantovani (ex Meta FAIR) and Lucas Tucker (ex Adobe). Our mission is to make AI more human by producing high quality audiovisual datasets for frontier labs. We are hiring a Forward Deployed Engineer to sit at the intersection of our platform and the real-world operations that power it. This is a hands-on, execution-focused role. You will own the end-to-end lifecycle of getting people onto our platform, procuring the data we need, and writing the scripts that turn collected footage into clean, usable training data. What you'll do Own data procurement: identify, recruit, and qualify participants for active data collection projects. Meet with researchers from top labs to build the highest quality video datasets for them. Write and maintain post-processing scripts that clean, validate, and package thousands of hours of video data for training world models. Identify and resolve operational issues — failed video uploads, payment problems, etc Build lightweight internal tools and automations to reduce manual work and increase throughput as we scale. Who you are A technically savvy generalist with impressive professional or academic accomplishments, comfortable writing data pipelines, making deals, and meeting with top researchers. Someone who gets things done Detail-oriented and reliable when managing people and data at scale, where errors have downstream consequences. Comfortable working in ambiguous, fast-moving environments Even better if you have Background in research operations, data labeling platforms, or data collection. Experience building large-scale data processing pipelines. Background in ML research. You will thrive in this role if You are energized by operational work that has immediate, visible impact. You treat broken processes as engineering problems worth solving properly. You hold yourself to a high bar for data quality, knowing this determines model performance.

About Us Velvet is a data research company building the datasets that power the next generation of multimodal AI. Founded by Lucas Mantovani (ex Meta FAIR) and Lucas Tucker (ex Adobe Infra), our mission is to make AI more human by producing high-quality audiovisual training data for frontier labs. We're hiring a Research Scientist to develop and fine-tune models for video and audio data processing and enhancement, as well as to conduct data-oriented research that pushes the boundaries of multimodal quality. What You'll Do Research, develop, and fine-tune models for audio and video enhancement — including denoising, super-resolution, speech restoration, and perceptual quality improvement — ensuring outputs meet the standards required for frontier model training. Experiment with novel architectures, training objectives, and data augmentation strategies to improve model performance across diverse and noisy real-world audiovisual data. Build evaluation frameworks and benchmarks to rigorously measure enhancement quality, guiding iterative model improvement. Collaborate with infrastructure and data pipeline engineers to integrate trained models into large-scale processing workflows that handle wide variation in speech, visual quality, and format. What We're Looking For Strong research background in deep learning, with hands-on experience training and fine-tuning models for audio processing, video processing, or related domains. Proficiency in PyTorch. Experience designing and running experiments at scale. Solid understanding of signal processing fundamentals and how they inform model design for enhancement tasks. A publication track record or demonstrated research output in relevant areas (audio/speech enhancement, video restoration, generative models, multimodal learning). Ability to work effectively in an early-stage environment where scope is broad and priorities shift fast. Even Better Prior work at a frontier AI lab or data company focused on multimodal data. Experience fine-tuning large pretrained models (diffusion models, autoencoders, or transformer-based architectures) for perceptual quality tasks. Familiarity with perceptual quality metrics and human evaluation methodologies for audio and video. Track record working with datasets spanning tens of thousands of hours of audio or video. You'll Thrive Here If You're excited by applied research with immediate, visible impact on data quality and downstream model performance. You move fluidly between reading papers, writing training loops, and analyzing failure cases. You hold yourself to a high bar for rigor — because you understand that model quality directly determines the value of the data we produce.

About Us Velvet is a data research company building the datasets that power the next generation of multimodal AI. Founded by Lucas Mantovani (ex Meta FAIR) and Lucas Tucker (ex Adobe Infrastructure), our mission is to make AI more human by producing high-quality audiovisual training data for frontier labs. We're hiring a Founding Machine Learning Engineer to build the pipelines that turn raw footage into clean, structured training data. This is a hands-on, execution-heavy role at the intersection of ML engineering and research. You'll own the full lifecycle — from writing and testing processing scripts to deploying them at scale across thousands of hours of video. What You'll Do Build and enhance post-processing pipelines that clean, validate, and package large volumes of video and audio data for multimodal model training. These pipelines must handle wide variation in speech, visual quality, and format — making robustness a huge engineering challenge. Deploy and fine-tune open-source models for speech recognition, speaker diarization, video segmentation, and related tasks. Design infrastructure for large-scale distributed processing — parallelizing thousands of compute jobs across cloud platforms and optimizing for throughput and cost. What We're Looking For Strong experience in ML infrastructure, speech/audio processing, or large-scale data pipelines. Proficiency in PyTorch. Familiarity with distributed job orchestration. Claude Code pilled. A bias toward shipping. You default to building, not theorizing. Ability to work effectively in an early-stage environment where scope is broad and priorities shift fast. Even Better Prior work at a data company or frontier AI lab. Track record building pipelines that process tens of thousands of hours of audio or video. Experience with infrastructure cost optimization or model fine-tuning for production use. You'll Thrive Here If You're energized by operational work with immediate, visible impact. You treat broken processes as engineering problems worth solving properly. You hold yourself to a high bar for data quality — because you understand it directly determines model performance.