Senior Machine Learning Engineer

Mission

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.

Outcomes - 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.

Skills - Core Technical Capabilities

Required:

• 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

Preferred:

• Experience with vector databases, embedding pipelines, and real-time retrieval systems
• Background in creative scoring, aesthetic modeling, or ad performance prediction

Competencies - Behaviors We Like to See

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