Machine Learning Operations (MLOps) Engineer

About the role

ABOUT GALLATIN

At Gallatin, we are rebuilding logistics infrastructure for the national security missions of the United States and allied partners. We build AI systems that determine how logistics decisions are made — not just how they're executed. From factory to foxhole, we operate at the layer where data becomes decisions, and decisions make the advantage.

WHAT YOU'LL DO

In this role, you will build the systems that move a model out of a notebook and into the hands of a planner. Sometimes, that means deploying to an air-gapped rack in a tent instead of a VPC. You will own the path from training run to deployed capability: the infrastructure it runs on, the release process that ships it, the evaluation harness that proves it works, and the telemetry that tells us when it stops working.

Our AI/ML team works across retrieval-grounded systems for doctrine and logistics data, document and feature extraction, military symbol recognition, optimization and movement models, and an LLM agent platform. This role underpins that work: building the infrastructure, release processes, and evaluation systems that make it shippable and keep it honest in production. You will have plenty of room to shape how we build it.

TRAINING & SERVING INFRASTRUCTURE

- Own model training, fine-tuning, and batch inference infrastructure across AWS (SageMaker, EKS) and on-premises GPU hardware.

- Stand up and tune LLM inference serving: vLLM-class stacks, quantization, continuous batching, KV-cache and throughput sizing. Make the hosted-vs-local call with numbers behind it.

- Build for DDIL: local inference with configurable fallback, degraded-mode behavior, and sane resource envelopes on hardware we do not get to choose.

RELEASE & REPRODUCIBILITY

- Build CI/CD for models and pipelines: versioned datasets, a model registry, promotion gates, and rollback that actually works under pressure.

- Own infrastructure as code, containerization, and GitOps deployment across environments ranging from a dev cluster to a disconnected enclave.

- Make reproducibility a hard requirement. Any result we put in front of a customer or evaluator must be reproducible from a commit and dataset version.

EVALUATION & OBSERVABILITY

- Build and own the evaluation harness: regression suites for retrieval and extraction pipelines, LLM-as-judge pipelines with measured judge-human agreement, and adversarial and held-out sets.

- Instrument production for drift, latency, cost, retrieval quality, and failure modes, including quiet ones such as a retrieval miss that produces a fluent but wrong answer.

- Make our metrics defensible to external test and evaluation reviewers. “We think it’s good” is not a deliverable.

DATA & PIPELINE OWNERSHIP

- Own ingestion, versioning, and lineage for logistics and doctrinal data drawn from a heterogeneous set of authoritative sources.