Machine Learning Engineer

━ About the Role

We are looking for an ML Engineer to own model development on a federal data and AI program: self-supervised representation learning over tabular and temporal administrative data, language-model-based information extraction from unstructured records, supervised classification on learned embeddings, and model explainability. The work is performed against large linked datasets inside an access-restricted analytic environment with approved tooling, no external connectivity, and CPU-only compute.

This is a modeling role, not a pipeline role. You will be accountable for architecture selection, training, calibration and validation of the models the program depends on, and for the quantitative judgment behind each of those decisions. Data engineering is part of the job; it is not the center of it.

━ What You Will Do

Representation Learning

• Design and train self-supervised tabular transformers (SAINT, TabTransformer, FT-Transformer) and temporal sequence models over longitudinal event data, with the outcome label withheld during pretraining.
• Select pretraining objectives, execute pilot runs, and validate embedding quality before committing to full-scale training.
• Benchmark candidate architectures against gradient-boosted baselines and document comparative results.

Language Model Extraction

• Deploy locally hosted open-weight language models for structured information extraction from unstructured text, including prompt design, constrained output schemas and post-processing.
• Quantize and tune models for CPU-only inference at volume without unacceptable degradation in extraction accuracy.
• Implement the dual-coding evaluation protocol — stratified sampling, independent human coding, field-level agreement measurement — and report accuracy before volume extraction.

Supervised Modeling and Calibration

• Train gradient-boosted classifiers (LightGBM, XGBoost) on learned embeddings under severe class imbalance.
• Perform threshold calibration on the validation partition with the cost of each error type made explicit, and report calibration curves alongside discrimination metrics.
• Detect and prevent leakage across temporal splits and across linked record sources.

Explainability

• Generate additive feature attributions (SHAP) at the individual-prediction level and aggregate them across cohorts and subgroups.
• Verify that attributions are stable across model retrains and consistent with domain expectations.

Supporting Data Work

• Maintain the feature and inference pipelines the models depend on (Spark, Parquet, Python), and version every dataset, model and experiment in MLflow so reported results regenerate.

━ Mathematical Foundation

This is a modeling position and we weight mathematical grounding heavily. Applied fluency is expected in the following; the interview will test it.

Area

Expected depth

Probability and statistical inference: Estimators, bias and variance, sampling distributions; correct interpretation of confidence intervals and calibration curves

Linear algebra and optimization: Gradient-based optimization, regularization, convexity and its failure in deep networks; attention understood as matrix operations

Information theory: Cross-entropy, KL divergence, and the loss functions behind self-supervised and contrastive objectives

Evaluation theory: Discrimination versus calibration; ROC and precision-recall behavior under class imbalance; threshold selection by expected cost

Causal reasoning: Confounding and leakage; recognizing when a feature encodes the outcome or a downstream consequence of it

━ Technical Stack

Layer

Tools

Tabular representation learning: SAINT, TabTransformer, FT-Transformer; self-supervised pretraining objectives (masked-feature reconstruction, contrastive)

Temporal and sequence modeling: Transformer encoders over ordered event sequences (BEHRT-style), temporal convolutional networks

Language models: Open-weight instruction-tuned models (Llama, Mistral, Qwen families) hosted locally; quantized CPU inference via llama.cpp or equivalent

Supervised learning: LightGBM, XGBoost, CatBoost; scikit-learn

Deep learning framework: PyTorch

Explainability: SHAP (TreeSHAP, KernelSHAP), integrated gradients

Data and MLOps: Spark, Parquet, pandas, NumPy; MLflow for experiment tracking and model versioning

Environment: Cloud-hosted, access-restricted analytic environments

━ What We're Looking For