Description
We're looking for a Principal ML Engineer to drive down cost per asset and drive up throughput across the ML pipelines behind Adobe's multimodal foundation models (image, video, audio). You'll own the end-to-end economics of turning billions of raw assets into training-ready data, making it faster and cheaper without sacrificing data quality.
Responsibilities
- Own cost per asset across the full pipeline, including compute (GPU and CPU), storage, data access and movement, and reprocessing, and identify the biggest levers to reduce it
- Optimize pipeline DAGs to cut cost, such as running cheap filters first so expensive model stages only process the assets that matter and reduce waste from unnecessary duplications
- Improve throughput of pipelines running specialized and general-purpose models (VLMs, LLMs, captioners, embedding and quality models) through batching, parallelism, quantization, and serving optimizations
- Distill larger models into smaller, cheaper ones that preserve labeling and filtering quality
- Decide where specialized models, general models, or cascades give the best cost/quality tradeoff
- Build evaluation frameworks to ensure efficiency gains don't degrade training data quality
- Develop a deep understanding of what makes data valuable for training LLMs, VLMs, and diffusion models, and encode it into filtering, scoring, and selection
- Partner with modeling teams to connect pipeline changes to training outcomes
- Work with data platform teams on orchestration (e.g., Ray), scheduling, and storage so ML workloads run cost-efficiently on shared infrastructure
Requirements
- 10+ years in ML systems, ML infrastructure, or applied ML engineering, including work at billions-of-assets scale
- Track record of cutting the cost of large-scale ML data pipelines across compute, storage, and data access (e.g., early filtering, deduplication, model cascades)
- Deep expertise in optimizing inference for VLMs, LLMs, or other large models, including batching, quantization, serving, and distillation
- Experience with data curation and data quality for training generative or multimodal models
- Strong Python, plus a systems-level language (C++, Rust, Go, or Java), and experience with distributed frameworks such as Ray or Spark
- Clear communication across modeling, infrastructure, and data teams
- Bachelor's, Master's, or Ph.D. in CS, Engineering, ML, or a related field