What this is
A conversational, multi-agent platform we designed and built that walks an applicant through assembling a complete petition packet for extraordinary-talent and national-interest immigration categories (O-1A, EB-1A, EB-1B, EB-2 NIW). The applicant interacts with a single paralegal-style conversational interface; everything else — orchestration, domain agents, knowledge retrieval — runs behind it. Final packets are reviewed by a human before submission.
We share architecture and a live walkthrough under NDA.
Why we built it
Petition assembly for these visa categories is one of the hardest places to do generative AI well. The criteria are statutory, the evidence requirements are precise, the cost of a hallucinated citation is a denied case, and the people doing the work today are expensive specialists whose time is poorly spent on intake and routine drafting. We wanted to demonstrate that a multi-agent system could compress the early- and middle-stage work without compromising defensibility.
What it does
- Conversation-first intake, eligibility screening, evidence collection, drafting, and packet assembly across the supported visa categories.
- Versioned legal knowledge bases — corpora are ingested through a reproducible pipeline and version-pinned, so a case is always answered against a known-good snapshot of the rules.
- Hybrid retrieval (lexical + semantic) over those private corpora.
- Every answer carries a source tier, confidence, and citation back to the source — there are no uncited assertions.
- An append-only audit log captures every event in the case file, so every automated decision is reviewable and reversible.
Built with
| Layer | Stack |
|---|---|
| Models | Leading frontier LLM (generation), open-source embedding models (retrieval) |
| Backend | Python, FastAPI |
| Database | PostgreSQL with vector + full-text extensions |
| Cache & messaging | Redis |
| Object storage | S3-compatible |
| Frontend | React + TypeScript + Vite + Tailwind + Shadcn/ui |
| Orchestration | Disciplined agent runtime; open-source agent libraries where they fit |
| Reproducibility | Docker Compose (clean clone → running stack) |
What this means for you
If you have a high-stakes domain workflow — immigration, regulatory filings, clinical-trial protocols, grant petitions, complex insurance claims, anything where the cost of a hallucinated answer is unacceptable — we can:
- Build a multi-agent conversational platform tailored to your domain, with the same source-tier / confidence / citation discipline.
- Deliver versioned knowledge bases with reproducible ingestion, so your domain rules can be updated, snapshotted, and rolled back like code.
- Engineer append-only audit trails so every automated decision is reviewable for regulators, reviewers, and your own QA.
- Ship the system as code your team owns, deployable on your cloud or on premises.
Want to see the platform in a live walkthrough under NDA? Contact us.