Data & ML Engineer

As a Data & ML Engineer you will build the data and model layer behind an AI-enabled decision-support system operating inside an accredited environment. That work covers ingestion from many source systems, resolution of incoming records against a shared data model, relevance scoring, and generation of explanations a user can act on and defend.

Three characteristics make this a substantial technical challenge. The incoming data is predominantly low-signal, which means a model can report strong overall accuracy while failing on the cases that matter most. Every output must remain traceable to the underlying sources, because a person downstream is accountable for the result. Record matching is probabilistic rather than exact, so false matches and missed matches both carry meaningful cost.

You will not be starting from an empty repository. We operate an established platform for source custody, extraction, and retrieval, and its architect is a member of this team, so existing design decisions are documented and accessible. Your work will focus on new capability rather than maintenance: record matching, calibrated scoring, and grounded generation, hardened for the target environment. We build with current tooling and expect the same, including the use of AI assistance in our own engineering practice.

This is a fully remote role with occasional travel (up to 25%) to DEFCON AI HQ, customer sites, and vendor facilities as required.

Key Responsibilities 

The technical work falls into four areas. Deep expertise in all four is not expected, so please indicate where your depth lies when you apply. The engineering standards that follow apply to everyone on the team.

Data Modeling and Record Matching

• Design and maintain the graph of entities, records, and the typed relationships between them
• Implement probabilistic matching, including blocking, candidate generation, pairwise scoring, clustering, and threshold policy
• Build deduplication and known-record suppression
• Establish provenance so that every node and edge traces to the source that asserted it
• Produce interface and data-flow design documentation detailed enough to serve as an implementation reference for other engineers

Scoring and Calibration

• Develop relevance and priority models over large, imperfect record sets
• Own calibration and threshold design, establishing what a score means rather than only how it ranks
• Design abstention policy that routes uncertain and high-risk cases to a person rather than returning a confident answer
• Perform feature engineering, establish baselines before introducing complex models, and conduct error analysis that accounts for the differing cost of false positives and false negatives

Retrieval and Generation

• Implement embeddings, vector storage, and retrieval across a large provenance-tracked evidence base
• Integrate language models through an approved managed service, and maintain a self-hosted or open-weight alternative within the same boundary
• Design prompts and output schemas
• Bind generated text to cited source records, and treat "insufficient evidence" as a valid system response rather than forcing a conclusion
• Own model packaging, serving, versioning, and rollback

Pipelines and Source Handling

• Build secure ingestion, transformation, validation, and publishing across structured, semi-structured, and unstructured sources
• Implement quality checks, schema validation, lineage capture, and audit logging
• Establish source drift detection so that degradation is surfaced rather than carried into the analysis
• Generate statistically representative synthetic data so that development can proceed ahead of live data access

Engineering Standards

• Work to the data model and standards set by the Data Lead, who approves designs and owns them through customer review
• Document assumptions, caveats, transformation logic, and known limitations, since deliverables are formally reviewed
• Instrument telemetry so that measurement does not require manual reconstruction
• Maintain the audit trail covering recommendations, human overrides, and model versions
• Submit model and pipeline changes through a gated release process rather than deploying in place

Required Qualifications

• 5+ years of experience in data engineering, data architecture, applied machine learning, ML engineering, or production analytics engineering
• Strong Python and SQL, with demonstrated experience working with large, imperfect operational data
• Experience delivering systems for sustained operational use rather than exploratory analysis alone
• Routine use of AI-assisted development, with informed judgment about where it adds value and where its output requires verification
• Ability to explain a technical decision to a stakeholder who must defend that decision without understanding its internals
• US Citizenship Required
• Active US Secret clearance. The work is performed in a controlled government cloud environment and requires a favorable investigation and CAC eligibility from the start
• Elevated personnel security requirements apply to portions of this work and are discussed during screening
• Willingness to travel up to 25% to customer sites, DEFCON AI HQ, and vendor facilities as required

Preferred Qualifications

• Clearance: active Top Secret
• Matching: direct experience applying probabilistic matching to inconsistent identity data, including names, dates, addresses, and identifiers, and familiarity with the failure modes of each. Record linkage, master data management, or identity management. Graph data modeling. PostgreSQL and pgvector or comparable. Graph algorithms applied in production
• Modeling: model calibration and threshold design. Cost-sensitive learning where error types carry unequal consequences. scikit-learn, XGBoost, PyTorch
• Retrieval and generation: retrieval-augmented generation in production. Prompt and output-schema design. Establishing that generated output remains grounded in its sources, and testing to confirm it. Self-hosted or open-weight model operation. Fine-tuning, adapters, or custom embeddings
• Pipelines: AWS Glue, Airflow, dbt, Spark, Kafka, or NiFi. Unstructured and semi-structured document ingestion. Synthetic or representative test data generation
• Environment: federal DevSecOps, RMF, ATO, or DoW cloud environments. Hardened base images. Experience advancing a pipeline from development through accreditation and deployment
• Domain: sensitive federal or defense data, and work performed under privacy or comparable handling constraints
• Responsible AI: documentation, model cards, fairness testing, and model monitoring. NIST AI RMF or comparable practice

What Success Looks Like

• A data model that the rest of the team builds on without needing to redesign it
• Matching decisions that can be explained and defended to a non-technical reviewer
• Models whose miss rate is characterized, not only their overall accuracy
• Generated explanations that assert no more than the sources support, with the citation path intact
• Pipelines that surface problems early and trace them to a specific source
• Consistent development progress, including during periods when live data is not yet available