Machine Learning Engineer, DEFCON AI

• 5+ years of experience, including a production or near-production retrieval-augmented (RAG) system you built yourself
• Ability to speak in detail to your retrieval design, which vector store you used and why, how you tested grounding, what citation failures looked like in practice, and how rollback worked
• Strong Python, with hands-on experience in embeddings and vector retrieval at scale
• Clarity on what actually shipped in past work — prototype, proposal, or deployed code — since that distinction matters more here than the title on a resume
• US Citizenship Required
• Active US Secret clearance required to start

Desirable

• As ML Engineer, Retrieval & Grounded Generation, you'll build embeddings, vector storage, and retrieval at scale, and integrate language models so that every piece of generated text is bound to cited source records and citation failures are tested for rather than assumed away
• You'll also own prompt and output-schema design; model packaging, versioning, serving, and rollback; and the telemetry hooks that make later measurement possible without manual reconstruction - real infrastructure for a real production system, not a demo
• Implement embeddings, vector storage, and retrieval across a large, provenance-tracked evidence base
• Integrate language models so generated text is bound to cited source records; test for citation failures rather than assuming them away
• Design prompts and output schemas
• Own model packaging, versioning, serving, and rollback
• Instrument telemetry for retrieval and generation quality, recommendation/version attribution, overrides, abstentions, grounding failures, latency, throughput, and measurement events defined with Model Test
• Provide bounded model assistance for difficult narrative extraction where deterministic processing is insufficient, with every output tied to its source passage
• Supply the recorded rule context to every model-assisted step, so each output carries the exact rule versions and ordered context it received
• Maintain a modular in-boundary serving path, self-hosted or managed, alongside the primary managed inference service, so the platform does not depend on one provider’s availability or approval
• What Success Looks Like:
• Generated explanations that assert no more than the sources support, with the citation path intact and citation failures tested rather than assumed away
• A retrieval system that performs at scale on a large, provenance-tracked evidence base
• Model rollback that works when it's needed, with telemetry complete enough that measurement does not require manual reconstruction