GDI ​/ Document Intelligence & NLP Specialist

Boston, United States | Posted on 05/20/2026

ATS+Partners is seeking a GDI / Document Intelligence & NLP Specialist who brings verifiable, multi-year experience implementing Intelligent Document Processing and Generative Document Intelligence solutions — ideally for government or regulated-industry clients —with documented extraction accuracy benchmarks from prior deployments. This role exists specifically to address the two most consequential gaps in the current proposal:

• Minimum Criterion 1 requires five or more years of experience implementing AI Automation and GDI solutions, preferably from municipal or government clients. The contractor’s own professional history must provide the concrete, dated implementation record that satisfies this threshold.
• Minimum Criterion 2 requires a proven ability to achieve 80%+ data extraction accuracy on unstructured documents. The contractor must be able to supply documented benchmark results from prior engagements — not theoretical capability claims — that an evaluator can verify.

The contractor will serve as the project’s extraction engine expert, responsible for the design, configuration, and continuous improvement of all AI document extraction pipelines. They will own prompt engineering for all four target workflows, configure the model ensemble’s consensus logic, build the field-level confidence scoring framework, and manage the human-in-the-loop feedback loop that drives continuous accuracy improvement throughout the three-year contract term.

• Design and configure AI extraction pipelines for all four target workflows:
  Invoice Processing, Vehicle Insurance Certificate Reconciliation, Health Insurance Reconciliation, and Audit Preparation.
• Develop and maintain custom prompt templates for GPT-4o and Claude AI, including field extraction schemas in JSON format, field-level instructions with positive and negative examples, chain-of-thought validation prompts, and conditional extraction rules.
• Configure the Lang Graph consensus orchestration layer: define how discrepancies between the two model outputs are resolved, what confidence thresholds trigger automatic pass-through vs. HITL routing, and how reviewer corrections are captured in graph state for retraining.
• Build and manage the document classification pipeline: design the taxonomy for audit document categorization, train or configure the classifier against City document samples, and validate classification accuracy.
• Establish and execute the accuracy benchmarking protocol: define hold-out test sets for each workflow, run pre-pilot benchmarks, pilot accuracy benchmarks, and quarterly production accuracy audits.
• Lead prompt regression testing: maintain a library of annotated sample documents that are re-run against every prompt update and model version change to detect accuracy regressions before they reach production.
• Manage the continuous learning loop: export reviewer corrections from the HITL queue, curate training data, coordinate model fine-tuning or prompt updates on a quarterly basis.
• Configure pre-processing pipeline parameters: image quality thresholds, deskewing settings, multi-page document segmentation rules, and PDF/A normalization parameters for each document type.
• Provide domain-specific extraction expertise for ACORD insurance certificates, government invoices, AP aging schedules, GL trial balances, and health insurance enrollment forms.
• Document all extraction schemas, prompt versions, confidence threshold configurations, and model accuracy results in a version-controlled Model Registry maintained throughout the contract.

• AI/GDI implementation experience: 5+ years implementing Intelligent Document Processing or Generative Document Intelligence solutions in production environments.
• Confidence scoring and HITL design: experience building tiered confidence routing logic connecting AI extraction outputs to human review queues.
• Documented extraction accuracy benchmarks: must be able to provide benchmark results (with methodology) from at least two prior IDP deployments showing 80%+ field-level extraction accuracy on unstructured documents.
• Government or…