Docusign and NVIDIA Built a 900M-Parameter Extractor That Reads Contract Tables 20x Faster
AI Engineer · youtube · 2026-09-16
On the AI Engineer podcast, Docusign's Hiral Shah and NVIDIA's Sean Sodha explain why agreements are an engineering problem.
- Roughly $2 trillion in negotiated value sits locked in agreements nobody revisits; Docusign serves 2M paying customers and 1M agreements per day
- Tables break generic extraction: pricing tiers, rate cards, SKUs and SLAs are exactly what users need and exactly what line-by-line reading destroys
- The pair built a 900M-parameter vision-language model designed as an extractor, not a generator — a single pass returns reading order, semantic structure and preserved tables, replacing a stack of layout models
- Lesson in restraint: the purpose-built model ran table extraction 20x faster than general alternatives, with lower latency and cost
- Q&A closes on whether OCR is going away
More from Research
- Radial gate objective cuts diffusion transformer FID from 16.24 to 11.83 — Junha Park · 2026-11-02
- REA: async brain-cerebellum runtime for robots where policies never wait on the LLM — Pengfei Zhou · 2026-11-02
- Microsoft's MemSteer turns unlabeled user histories into agent memory supervision — MicrosoftResearch · 2026-11-02
- SciGame benchmark uses counterfactual worlds to expose LLM agents' priors vs evidence — Yusheng Zhao · 2026-11-02
- On Terence Tao's blog, Frank Calegari calls on mathematicians to adapt after OpenAI's 700-paper release — Hesamation · 2026-11-02
- Tsinghua & Microsoft's HarnessMM cuts agent token use by up to 87.93% while raising success rate — rohanpaul_ai · 2026-11-02