Production system
SkipLegal: multi-agent H-1B document intelligence
Production system for Crossing, Data Science GQP MVP
- Timeline
- Aug 2025 to Dec 2025
- Affiliation
- Crossing x WPI industry-sponsored Graduate Qualifying Project
- Team
- Mansi Rupesh Mantri, Nandini Priya Sripada, Pradyumn Tendulkar
- Recognition
- Data Science GQP MVP, December 2025
- Review time
- 3 to 4 days down to 5 to 10 min
- Structural accuracy
- 95 to 99%
- Missing-document detection
- 99 to 100%
- Rules across fields
- 50+ over 45+
- Hours freed per firm per season
- about 2,000
- Cost per case
- 50 to 70% lower
What it is
A four-agent LLM pipeline on AWS that turns messy multi-format H-1B petition packages into schema-validated JSON and a single attorney-ready QC report, cutting review from 3 to 4 days to about 5 to 10 minutes.
Problem
Immigration attorneys spent days per petition manually checking Form I-129 packages and supporting evidence for missing documents, inconsistent fields and structural errors.
Approach
Agent 1 (intake and vision) collapses PDFs into a unified schema with Textract OCR; Agent 2A canonicalizes I-129 fields through a deterministic logic engine; Agent 2B runs LLM semantic validation on supporting evidence; Agent 2C is a cross-document comparator with an Amazon Bedrock RAG fallback for cases the rules cannot resolve. Orchestrated with LangGraph on S3, Lambda, Step Functions and EventBridge.
Highlights
Owned client-facing delivery: requirements discovery with Crossing engineers, acceptance tests, weekly demos, integration and handoff.
Built the computer-vision grouping, API schema and the Pydantic-validated canonical JSON contract consumed downstream.
RAG evaluation loops with RAGAS (faithfulness, context recall) and threshold-based regression gating across prompt and retriever changes.
Presented to eleven industry judges at the WPI GRIE Showcase in April 2026; recognized with the Data Science GQP MVP award.