Applied ML Engineer

About this role

Deepgram is seeking an Applied ML Engineer to bridge research and production by owning the end-to-end pipeline that transforms speech AI models from experimental checkpoints into scaled, reliable services. You'll collaborate with research scientists to productionize models, build deployment infrastructure, and establish quality gates across a hybrid GPU and cloud environment.

What you'll do

• Own the research-to-production pipeline, converting research checkpoints into production-ready models with repeatable processes
• Partner with research scientists to translate experimental code into robust, reproducible training and evaluation workflows
• Build tooling and abstractions that streamline model movement through training, evaluation, packaging, and deployment
• Design and maintain model release gates with automated evaluation, regression detection, and performance checks
• Optimize models for production inference including batching, memory tuning, and latency optimization
• Instrument production behavior and establish feedback loops to inform future model iterations

What they're looking for

• ML productionization and deployment pipelines
• Python and ML frameworks (TensorFlow, PyTorch, or similar)
• Infrastructure and DevOps (GPU compute, cloud platforms, containerization)
• Model optimization and profiling for inference
• Benchmarking, testing, and quality assurance automation
• Systems engineering and performance tuning
• Scripting and tooling development
• Collaborative problem-solving with research teams

Benefits

• Remote work, USA-based
• Work on cutting-edge speech AI at billion-scale
• Hybrid infrastructure environment (custom GPU datacenters + cloud)
• Opportunity to define ML delivery practices
• AI-first culture with emphasis on innovation
• Series C-backed company with strong investor backing

Likely interview questions

• Describe a time you took a research model or prototype and prepared it for production—what were the biggest bottlenecks?
• How do you approach optimizing a neural network for low-latency inference at scale?