MLOps Engineer

Builds the production substrate for AI systems: model gateways, prompt and tool registries, evaluation harnesses, observability stacks, and the cost-control instrumentation that keeps generative-AI economics viable at portfolio scale.

Treats AI infrastructure as platform engineering with a probabilistic execution layer, not as a separate discipline. Insists on the same SLO and on-call discipline for AI systems as for any other production service. Pushes back on per-team ad-hoc model integration in favor of shared platform primitives that scale with the portfolio.

• Model gateway design (LiteLLM, Bedrock, Vertex, custom)
• Provider abstraction and multi-model routing
• Prompt registry design with versioning and rollback
• Tool registry design and shared tool libraries
• Evaluation harness construction (offline + online)
• CI/CD for AI artifacts (prompts, tools, model configs)
• Token-level cost telemetry and per-request economics
• LLM observability platforms (Langfuse, Helicone, Arize, OpenTelemetry-based custom)
• Drift detection on inputs and outputs of production AI systems
• Latency and throughput SLO design for AI services
• Feedback loop construction (user thumbs, structured feedback, downstream business outcomes)
• Caching strategies for LLM responses
• Rate limiting at multiple granularities (user, session, tool, model)
• Cost attribution and chargeback / showback for AI workloads
• Commitment-tier optimization for hyperscaler AI services
• Incident-response process for AI-specific failure modes
• Container orchestration (Kubernetes, ECS) for model-serving workloads
• Infrastructure-as-code for AI platforms

• Model gateway architecture with routing, retries, fallback, and cost attribution
• Prompt and tool registry design with versioning and environment promotion
• Evaluation harnesses integrated into CI/CD for AI artifacts
• LLM observability stacks (token-level cost, latency, drift, abuse)
• Feedback collection pipelines closing the loop between users and model owners
• SLO design for non-deterministic systems
• FinOps for generative-AI workloads

• Twelve to twenty-four week AI platform builds for clients scaling pilots to portfolio
• Embedded MLOps augmentation for client AI platform teams
• Cost-discipline engagements when generative-AI bills exceed planned budgets
• Reliability and observability engagements for clients with production AI incidents
• Migration engagements consolidating per-team AI integrations onto a shared platform