Senior Machine Learning Engineer, Kargo

• Techies who want to build the future
• Creatives who want to design it better
• Communicators to win business
• Collaborators to build it
• Data pros who turn numbers into insights
• Product builders who turn ideas into innovations
• Anyone eager to be on a team that doesn’t stop to ask what’s next, because they’re already building it
• 5+ years in ML engineering or MLOps, with shipped production systems involving LLMs, VLMs, or multimodal architectures
• Expert in Python and PyTorch (or TensorFlow), plus distributed training frameworks (Ray, PyTorch Lightning, Horovod)
• Hands-on with MLOps tooling: MLflow, Weights & Biases, Kubeflow, Argo, or Airflow for orchestration, experiment tracking, and automated retraining
• Cloud-native ML deployment on AWS (SageMaker), GCP (Vertex AI), or Azure ML, with infrastructure-as-code (Terraform, Helm)
• Production fluency with Docker, Kubernetes, and CI/CD patterns for ML
• Strong SQL, data pipeline, and feature store design for scalable experimentation

Desirable

• Own the evolution of Finetouch, Kargo's creative scoring system, by leading the design and production deployment of multimodal ML models that quantify creative quality and predict ad performance
• This role is the technical anchor for the Creative Sciences Platform — translating research in LLMs, VLMs, and multimodal learning into scalable, reliable systems that creative and product teams build on
• Success means Finetouch becomes faster, smarter, and more trusted as the intelligence layer behind Kargo's creative analytics
• What Success Looks like in 6-12 months
• Ship the next generation of Finetouch. Deliver a measurably improved version of the creative scoring model — better predictive accuracy on creative performance, expanded multimodal signal coverage (visual + text + engagement), and validated lift over the current baseline
• Stand up production-grade MLOps for Creative Sciences. Establish end-to-end pipelines (training, fine-tuning, deployment, monitoring) on MLflow/Kubeflow/Ray Train so model iterations move from notebook to production in days, not weeks, with full reproducibility
• Scale distributed training and inference. Reduce training time and inference cost on multimodal/VLM workloads through Ray, PyTorch Distributed, and right-sized cloud infrastructure — enabling larger models and faster experimentation cycles
• Expose Finetouch as a platform. Build and operate the APIs, embedding services, and model endpoints that let Glossi and other Kargo creative platforms consume scoring in real time, with documented SLAs and integration patterns
• Operationalize model reliability. Deploy real-time monitoring, drift detection, and alerting so production model degradation is caught before it affects creative decisions, with clear runbooks and on-call ownership
• Research-to-Production Judgment
• Knows when a model is good enough to ship vs. when it needs another iteration — doesn't over-engineer or under-validate
• Translates papers and prototypes into systems that survive production traffic, monitoring, and on-call
• Systems Thinking at Scale
• Designs for the second and third version of the model, not just the first — pipelines, abstractions, and infra that compound over time
• Optimizes the full stack: training cost, inference latency, and developer iteration speed, not just model accuracy
• Cross-Functional Translation
• Explains multimodal modeling tradeoffs to Product and Creative stakeholders in terms of business impact, not architecture diagrams
• Partners with Data Science and Platform Engineering as co-owners, not handoff points
• Operational Ownership
• Treats drift, latency regressions, and silent failures as personal — instruments systems so problems are caught early and root-caused fast
• Documents architecture, decisions, and runbooks so the platform outlives any single contributor