Engineering

Running a frontier model on every document doesn't scale at 1M+ documents a day. In part two of our series on how Docusign processes agreements at scale, we cover how we trained a smaller, faster model to match — and in production, beat — the accuracy of the larger model it replaced, at a fraction of the cost.

How did Docusign take AI coding adoption from single digits to near-universal in about a year? Not with a top-down mandate, but with a peer-led program designed to establish guardrails and embed AI coding into individual development teams. We detail our journey, from an early pilot to a grassroots multiplier network to a suite of internal AI agents now automating code review and repetitive engineering work.

Docusign's AI engine processes over a million agreements a day, extracting more than 50 data points per document at sub-cent inference cost. This is the first post in a four-part series on how we built that pipeline, starting with the orchestration architecture that routes every document through pre-processing, model inference, and validation at enterprise scale.

Contract tables — the SLA terms, rate cards, and pricing schedules buried in enterprise agreements — are among the hardest data to extract accurately at scale. Docusign has integrated NVIDIA's Nemotron Parse, a vision-language model purpose-built for document understanding, into its document processing pipeline to solve this. Table extraction powered by Nemotron Parse is now accepting beta customers with full general availability ahead.

Every mature codebase accumulates dead code: expired feature flags, unreachable branches, logic that shipped years ago and never got cleaned up. So we built Vulture to automate cleanup. It's already removed ~150,000 lines of dead code from our legacy eSignature codebase in one year. Here's how we built it.

Manual field mapping was one of the most common sources of user-reported errors in Workflow Builder workflows. Traditional semantic matching couldn't help because most eSign templates lack meaningful field labels. So we built a two-stage AI pipeline that identifies what each field represents by analyzing the document visually, then auto-maps those fields to Web Form data, cutting configuration time from 30 minutes to 20 seconds.