An LLM-Native Forms Platform

Rather than bolting LLM features onto an existing forms product, Forms Lab was designed from the ground up around LLM capabilities.

The Core Insight

Government forms are structured documents. They encode a data collection specification — what information to gather, in what format, under what conditions. An LLM can extract this structure directly from a PDF.

Architecture

The system separates three concerns:

• DataCollectionSpec — What to collect: fields, types, constraints, conditions, sensitivity classifications
• FormSpec — How to present it: pages, sections, delivery modes, visual layout
• Submission — Collected data, linked to the exact spec version

This separation means the LLM extraction produces a DataCollectionSpec, and the system can generate multiple presentation formats from the same underlying data model.

What Makes This Different

• Two LLM integration points, not one: structured extraction (PDF → spec) and constrained command generation (intent → edit batch). Both use different techniques — free-form structured output vs. tool-use — for different problems.
• Not just chat: The LLM does structured extraction and constrained command generation, not open-ended conversational interaction
• Systematic evaluation: Every extraction is scored against ground truth with quantitative metrics
• Production-grade: Full deployment pipeline, not a notebook demo
• Standards-compliant: Output meets USWDS accessibility and design standards

See: Architecture | Data Model | Design Decisions