Your Agreements Are a Database You Can't Query — Hiral Shah, Docusign & Sean Sodha, NVIDIA
Summary
The session addresses the massive engineering challenge of extracting structured, queryable data from enterprise agreements, which are often trapped in unstructured formats like PDFs. The scale is immense: Docusign processes approximately one million agreements daily, representing a potential $2 trillion in locked-up negotiated value. To solve the critical problem of complex tables (pricing tiers, SKUs, rate cards) that break generic extraction tools, DocuSign partnered with NVIDIA. They developed Neotron Parse, a purpose-built, small Vision Language Model (VLM) designed specifically as an extractor rather than a generator. This model provides single-shot processing for layout understanding, semantic structure, and accurate table preservation, achieving significantly higher speed and lower latency compared to general-purpose alternatives.
Key takeaways
-
Scale of the Problem
3:40
The enterprise agreement data corpus is massive, involving 1.9 million paying customers and a billion users, requiring the structuring of approximately one million agreements per day. The difficulty lies in the fact that agreements are hierarchical, not flat, and contain vital terms (pricing, SLAs) that are difficult to query.
-
The Failure of Generic Extraction
5:59
Traditional document extraction tools or generic VLMs fail when encountering complex tables, merged cells, or nested columns, leading to significant operational overhead for legal and procurement teams.
-
Efficiency of Purpose-Built Models
9:40
The Neotron Parse model, a small VLM (~850-900 million parameters), was designed for extraction, not generation. It demonstrated superior performance, running table extraction approximately 20 times faster than general alternatives, leading to lower latency and cost at scale.
Technical details
-
Neotron Architecture
470s
The solution utilizes a multi-stage process: Neotron Retriever handles finding the correct document/page from a large corpus, and Neotron Parse handles the complex extraction. Neotron Parse is a single-shot VLM designed to output semantic formatting, reading order, and preserved table structure.
-
Model Optimization and Deployment
750s
The model is optimized for efficiency, with paths toward deployment on FP8 and VFP4. The architecture supports advanced techniques like quantization and multi-token generation, ensuring high performance while maintaining accuracy for enterprise use cases.
-
Data Processing Pipeline
630s
The system transforms agreements into actionable insights by automatically structuring metadata and complex pricing data into downloadable formats (e.g., CSV) and making them available via API, creating a proprietary organizational data model.
Mentioned resources
- Docusign
- NVIDIA
- Neotron Parse
- NVIDIA NIM
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.