Senior Model Evaluation Analyst

Ignite IT•Suitland, MD

About The Position

Ignite IT Hub is seeking a senior Model Evaluation Analyst to assess the accuracy and reliability of the U.S. Census Bureau's LLM Autocoder. You will maintain evaluation datasets, run comparisons, investigate coding errors, and produce evidence for Government review. Working with the contract’s Data Scientist and Government subject matter experts (SMEs), you will evaluate enhancements for the American Community Survey (ACS) and expansion to other surveys and fields.

Requirements

Bachelor's degree in statistics, data science, mathematics, computer science, or a related field, or equivalent relevant professional experience.
8+ years in quantitative analysis, data quality, or model validation, including 3+ years evaluating ML, NLP, or automated classification systems.
Strong Python and SQL skills for preparing datasets, automating analyses, calculating metrics, and producing reproducible reports.
Practical understanding of sampling, class imbalance, reference-label quality, data leakage, confusion matrices, and uncertainty in performance estimates.
Ability to investigate discrepancies independently, document evidence precisely, and explain findings to technical staff and business or domain experts.

Nice To Haves

Experience evaluating LLM outputs, survey write-in coding, or detailed industry, occupation, and demographic classifications.
Experience with annotation or adjudication workflows, federal statistical data, and controlled environments for sensitive information.

Responsibilities

Prepare and maintain versioned evaluation datasets, reference labels, and test cases with documented provenance and representative coverage of coding categories.
Execute evaluation methods established with the Senior Data Scientist and calculate accuracy, joint coding rates, coverage, and results against agreed error thresholds.
Compare candidate models and coding methods with approved baselines; identify material category-level regressions that aggregate results may conceal.
Analyze difficult, ambiguous, and underrepresented cases; assemble examples for SME adjudication and track decisions and label corrections.
Develop repeatable Python or SQL analyses and evaluation checks, preserving the data, code, prompt, model, and configuration versions behind each result.
Maintain test evidence, review-comment disposition, and traceability from requirements and acceptance criteria to evaluated results and deliverables.
Prepare clear performance reports, tables, and briefings that explain findings and limitations; support pilot assessments, release reviews, and ongoing monitoring.