Document AI Engineer

Utility bills are among the hardest business documents to parse: multi-page layouts, riders spread over several tables, handwritten corrections and scans from fax machines. You will build and measure the extraction system that turns them into line items tied to a meter and a billing period. You will choose where to use layout models, rules and human review, define accuracy targets per field, and run the evaluation sets that tell us when a change helps. Customer bill data is never used to train outside models, so you will work with models and tooling that respect that constraint.

What you will do

• Own extraction accuracy for bill header fields, charge lines, usage, demand and rate schedule codes.
• Build evaluation sets, labeling tools and dashboards that show accuracy by utility and layout.
• Develop template-free and template-assisted extraction for new bill layouts.
• Design review queues that send low-confidence fields to analysts and learn from corrections.
• Work with tariff analysts to check extraction against their expected values.
• Keep latency and cost per bill within targets.

What you bring

• Four or more years building document understanding, OCR or information extraction systems in production.
• Strong Python and practical experience with PDF handling, OCR engines and layout models.
• Experience designing evaluation methods and metrics, not only models.
• Ability to debug ambiguous failures from raw documents to final output.
• Experience with confidence scoring and human-in-the-loop workflows.

Nice to have

• Experience with invoices, bank statements or other financial documents.
• Experience deploying models under data residency or no-training restrictions.
• Published work or open source contributions in document AI.