JOBSEARCHER

MLOps Engineer

JD – MLOps Engineer, Agentic AI Platforms
Role overview
Location: Dallas, TX | Type: [Full-time/W2] | Experience: 6–10 years
We need an MLOps Engineer who started in data engineering and now builds the platforms that run ML models and AI agents in production. You will own the path from raw data to deployed, monitored models and agentic workflows: feature pipelines, training automation, CI/CD, agent orchestration, MCP-based tool integration, serving, and observability. You will work with data scientists, AI engineers, and platform teams to make model and agent delivery repeatable, governed, and fast.
Key responsibilities: Design and maintain end-to-end ML pipelines covering ingestion, feature engineering, training, validation, deployment, and retraining. Build and run feature stores and training datasets with versioning, lineage, and reproducibility. Implement CI/CD/CT for models and agents: automated testing, model and prompt registries, approval gates, and staged rollouts (shadow, canary, A/B).Deploy models for batch and real-time inference on containerized, autoscaling infrastructure. Build, deploy, and operate agentic AI workflows (single and multi-agent) using LangGraph, Google ADK or Vertex AI Agent Builder. Design and run the agentic platform layer: agent runtime, orchestration, memory, tool registry, and shared services reusable across teams. Build and manage MCP (Model Context Protocol) servers and connectors that securely expose enterprise data, APIs, and tools (BigQuery, EDL, ticketing, internal services) to agents. Implement RAG pipelines: embedding generation, vector store management, retrieval tuning, and index refresh tied to upstream data pipelines. Set up agent evaluation and guardrails: output quality evals, grounding and hallucination checks, tool-call failure handling, human-in-the-loop approvals, and red-teaming. Monitor models and agents for performance, drift, latency, token usage, cost, and tool-call success, with tracing, alerting, and retraining or rollback triggers. Build scalable batch and streaming data pipelines that feed ML and agent workloads, with data quality checks built in. Manage infrastructure as code for ML and agent environments across dev, test, and prod. Enforce governance: least-privilege access for agent tools, audit trails of agent actions, explainability artifacts, and compliance requirements. Help data scientists and AI engineers move from notebooks and prototypes to production-grade, modular code.
Required skills: 4+ years in data engineering (ETL/ELT, data modeling, pipeline orchestration) and 2+ years in MLOps, ML platform, or LLMOps work. Hands-on experience shipping at least one LLM agent or agentic workflow to production. Agent frameworks: LangGraph, Google ADK, AutoGen, or Vertex AI Agent Builder.
MCP: building or integrating MCP servers and clients; tool/function calling and API integration for agents. LLM stack: prompt and context design, RAG, vector databases (Vertex Vector Search, pgvector, Pinecone, or Weaviate), and embedding models. Agent evaluation and observability: LangSmith, Langfuse, Arize Phoenix, Vertex AI evaluation, or equivalent. Strong Python and SQL; working knowledge of Scala or Java is a plus. Spark/PySpark and distributed data processing at scale. Orchestration: Airflow, Kubeflow Pipelines, or similar. ML lifecycle tooling: MLflow, Vertex AI, SageMaker, or Azure ML (model registry, experiment tracking).Containers and orchestration: Docker and Kubernetes (GKE preferred).CI/CD: GitHub Actions, GitLab CI, Jenkins, or Cloud Build. Infrastructure as code: Terraform or equivalent. Cloud: GCP preferred (BigQuery, Dataflow, Dataproc, GCS, Vertex AI, Gemini); AWS or Azure equivalents accepted. Monitoring and observability: Prometheus, Grafana, Evidently, OpenTelemetry, or cloud-native tooling. Agentic engineering competencies These are the skills that separate a production agent builder from an MLOps engineer who has only served models. Agent design: choosing between a single agent, a fixed workflow, or multi-agent; defining clear agent roles, stop conditions, and escalation paths. Context engineering: managing what goes into the context window (instructions, retrieved data, tool results, history), summarization, and prompt caching. Tool design: writing clean tool schemas and descriptions, structured outputs (JSON schema), idempotent tools, retries, timeouts, and safe fallbacks. State and memory: short-term and long-term memory, session state, checkpointing, and resumable long-running workflows (LangGraph checkpoints, Temporal, or Cloud Workflows).API and integration engineering: FastAPI or similar services, REST/gRPC, event-driven triggers (Pub/Sub), and async Python. Security and identity for agents: OAuth and service-account scoping per tool, secrets handling, prompt-injection and data-exfiltration defenses, and approval gates for write actions. Evaluation discipline: golden test sets, LLM-as-judge scoring, regression evals in CI, trajectory (step-by-step) evaluation, and simulation of edge cases before release. Cost and latency engineering: model routing (small vs large models), token budgets, caching, batching, and parallel tool calls. Agent observability: end-to-end traces of each run (prompts, tool calls, decisions) using OpenTelemetry GenAI conventions or LangSmith/Langfuse. Responsible AI and compliance: PII handling, data residency, audit logs of agent decisions, and alignment with enterprise AI policy. Human-in-the-loop design: where agents hand off to people, how reviewers approve or correct, and how that feedback improves the agent. Model adaptation (basic): when to prompt, when to use RAG, and when fine-tuning or distillation is justified. Business framing: scoping use cases, defining success metrics (accuracy, time saved, cost per task), and explaining agent behavior to non-technical stakeholders.
Nice to have: Multi-agent design patterns: supervisor/worker, planner/executor, and agent-to-agent (A2A) protocols. Model gateways and routing: LiteLLM, Apigee, or similar, with cost and rate-limit controls across LLM providers. Streaming: Kafka, Pub/Sub, NiFi, or Flink. Feature stores: Feast, Vertex Feature Store, or Tecton. Model serving: KServe, Seldon, BentoML, Triton, or vLLM.AI security: prompt-injection defense, PII redaction, and secrets management for agent tools. Experience in telecom, financial services, or other regulated, high-volume environments.
Certifications: GCP Professional ML Engineer or Data Engineer, AWS ML Specialty.
What success looks like:
  • (first 6 months)
  • Model and agent deployment time reduced from weeks to days through automated pipelines.
  • Every production model and agent has versioned data, code, prompts, and artifacts, plus drift and quality monitoring.
  • A shared agentic platform is live with a reusable MCP tool catalog, adopted by at least two teams.
  • Agent evaluation and guardrail checks run automatically before every release.
Education: Bachelor's in Computer Science, Engineering, or a related field, or equivalent experience.