PathPort

Live Demo!

pathport-final.vercel.app

Inspiration

The U.S. visa process is overwhelming. Millions of people who are fully eligible get denied every year not because they don't qualify, but because they are unprepared. One inconsistent answer and years of plans fall apart in a five-minute conversation. We built PathPort because nobody should fail a visa interview just because they didn't know what was coming.

What it does

PathPort helps you through the U.S. visa application process. An intake agent has a natural conversation with you to understand your situation, then generates a personalized document checklist based on your case. You upload your documents and the agent parses them, flagging gaps and inconsistencies. Finally, an adversarial multi-agent system, a prosecutor that finds every reason to deny your visa, and a defense agent that coaches you to close the holes, runs autonomously before putting you in a live video mock interview with a consular officer agent focused on your specific vulnerabilities.

Ethics!

Immigration is one of the highest-stakes areas to build in. A hallucination isn't a minor bug; it's potentially telling someone to say the wrong thing under oath to a federal officer. We thought carefully about this.

The most concrete thing we did was ground the adversarial debate in actual law using a RAG pipeline. Our RAG layer contains real source material, specific INA sections, BIA precedents, and USCIS Policy Manual provisions, and the prosecutor and defense agents are required to cite the specific provisions they're reasoning from by name. This anchors the analysis to real law rather than whatever the model happens to believe.

The defense agent also has an explicit honesty guardrail: if a point in your case is weak, it's required to say so rather than spin it. And throughout the product, PathPort never predicts outcomes or gives legal advice. It prepares you. The landing page states this explicitly: "consult a qualified immigration attorney for your case."

We also had plans to implement a swarm of verifier agents that independently cross-check any information before it reaches the user, catching contradictions or unsupported claims before they cause harm. We didn't have time to get there, but it's the first thing we'd build if we kept going.

How we built it

We built it on Next.js with TypeScript and Tailwind, and basically the entire AI layer runs through Anthropic's Claude API. We're using Claude for everything, the intake chatbot, document analysis via Claude's vision model, the adversarial debate between three AI agents, and the mock interview officer that scores your answers live. For the interview simulation specifically we layered in the Web Speech API for voice and face-api.js running in the browser for gaze tracking. The RAG system is custom-built; we wrote a local knowledge base of real immigration statutes and BIA case law, score chunks against the applicant's profile at runtime, and inject the relevant legal text straight into the prosecutor's system prompt so its arguments cite actual law. To protect privacy, all data persists in the browser's cache locally.

Challenges we ran into

Claude API costs hit us fast, running a prosecutor agent, a defense agent, and an evaluator on every interview loop adds up quickly, and we burned through credits faster than expected. We had to be aggressive about batching calls and trimming context to keep it under control.

The other one was the demo. To show the full pipeline working end to end, we needed a realistic applicant profile with real-looking documents; so one of us spent a solid chunk of the hackathon filling out fake I-20s, fabricating bank statements, and writing up a fictional Stanford admit letter for a made-up Korean MS student. Very fun use of our time :D

Accomplishments that we're proud of

The live interview is what we're most proud of. We record you live on camera, talking out loud to a synthesized consular officer voice, while an evaluator scores every answer in real time and breaks the moment you say something that would raise a red flag. Speech-to-text, face-tracking, and text-to-speech all running together in the browser.

The adversarial debate is the other one. A prosecutor agent reads your entire profile and documents and autonomously builds the strongest possible case for denying your visa; surfacing real vulnerabilities without being told what to look for, grounded in actual INA sections and BIA precedents. A defense agent fires back. They go two full rounds, then a judge agent delivers a final ruling on each vulnerability, so by the time the consular officer asks you a hard question, you've already heard the worst version of it.

And, everything actually works!

Built With

• claude