ReRoute

Inspiration

In early 2026, the TSA staffing crisis left thousands of travelers stranded at airports across the country. Flights were canceled with no warning, security lines stretched for hours, and rebooking systems buckled under demand. The coverage focused on frustrated business travelers and vacation plans gone wrong. But the people hit hardest were barely mentioned: the home health aide who needed to get to her Monday shift in Dallas, the construction worker who would lose a week of wages if he missed his Monday morning site call, the seasonal farm worker whose entire travel budget was a single $180 ticket.

White collar workers rebook with Amex points and airline lounge agents. They have employer travel desks, credit card trip insurance, and the financial cushion to absorb a $300 surprise. Blue collar workers have none of that. A canceled flight does not just mean inconvenience. It means lost wages, no-call no-shows, and sometimes losing a job entirely. And when they try to navigate their rights, they face airline phone trees, confusing DOT regulations, and rebooking portals that assume you have a laptop and time to spare.

We built ReRoute because we believe every worker deserves the same quality of travel advocacy that a Fortune 500 employee gets from their corporate travel desk. An AI agent that monitors your trip, knows the law better than the airline's own staff, and fights to get you where you need to be at the lowest possible cost, across every mode of transportation available.

What it does

ReRoute is a voice-first, multilingual AI travel recovery agent that acts as a personal advocate for workers facing flight disruptions. Instead of navigating airline phone trees or confusing rebooking portals, workers open ReRoute on their phone and talk to an agent that already knows their trip, their rights, and their options.

Core capabilities include:

Proactive trip monitoring — ReRoute watches your flight and detects disruptions before you even get to the airport. When something goes wrong, it immediately starts working on alternatives instead of waiting for you to notice.

Multi-modal rebooking — Unlike airline rebooking tools that only show you other flights on the same carrier, ReRoute searches across every transportation option: other airlines, Greyhound, Amtrak, FlixBus, and rideshare estimates. It compares them side by side with price, travel time, and arrival time so you can pick what works for your schedule and your budget.

AI-powered legal rights engine — ReRoute uses multi-step reasoning to analyze your specific disruption scenario against DOT regulations (14 CFR Part 259, 14 CFR Part 250) and calculates your exact compensation entitlement. It does not give you generic advice. It tells you: "You are owed $650. Here is why. Here is how to claim it."

Voice-first, multilingual interface — Many blue collar workers are more comfortable speaking than typing, and many speak Spanish as their primary language. ReRoute supports voice input via Web Speech API and operates fully in English and Spanish.

Disruption analytics dashboard — A data layer showing real-time TSA impact statistics, cancellation rates by airport, and cost impact on workers, providing transparency into the scale of the problem ReRoute is solving.

How we built it

ReRoute is a Next.js 14 application with a multi-model AI orchestration layer, all routed through a single unified API gateway.

Frontend (Next.js 14 + TypeScript + Tailwind + shadcn/ui): We built a mobile-first progressive web app designed to feel native on a phone. The entire UI is optimized for large touch targets (minimum 44px), high contrast text, and simple navigation. We used Framer Motion for page transitions and status change animations that give the app a polished, responsive feel. The design language is deliberately sturdy and trustworthy rather than sleek and minimal because our users are working people who need clarity, not aesthetics.

Multi-Model AI Orchestration via Lava Gateway: The core architectural decision was routing all AI calls through Lava's unified gateway, orchestrating three specialized models for different parts of the pipeline:

1. Hermes 3 (Nous Research, 405B) handles real-time message triage and intent classification. Every user message first passes through Hermes, which classifies it by type (rebooking, rights, status, general) and urgency level. This takes milliseconds and determines which downstream model handles the heavy lifting.

2. K2 Think V2 (MBZUAI, 70B) is our legal reasoning engine. When a user's message involves compensation or rights questions, K2 performs multi-step reasoning through DOT regulations, calculating exact dollar amounts owed and citing specific federal rules. We chose K2 specifically because compensation analysis is a genuinely complex reasoning task: you need to consider disruption type, delay duration, domestic vs international routing, voluntary vs involuntary bumping, and notification timing. K2's reasoning chain is displayed to users so they can see exactly how their compensation was calculated.

3. Claude Sonnet (Anthropic) generates the final user-facing conversational response. It receives the Hermes triage classification and K2 reasoning (when applicable) as context and produces warm, clear, jargon-free responses. It handles the Spanish language toggle and adapts its communication style for accessibility.

This three-model pipeline runs through a single Lava API endpoint, demonstrating that multi-model orchestration does not require multiple API integrations or complex infrastructure.

Analytics Layer (Hex): We integrated the Hex API to power a disruption analytics dashboard showing cancellation rates by airport, rebooking cost trends, and TSA impact statistics. This gives the app a data-driven backbone and shows the systemic scale of the problem.

Voice Input (Web Speech API): We implemented browser-native speech recognition with a visual pulsing indicator during recording. Transcribed speech is sent directly to the chat agent, making the entire experience usable without typing a single character.

Progressive Web App: ReRoute is installable as a PWA with a manifest.json, custom icons, and theme colors. Workers can add it to their home screen and it launches like a native app.

Challenges we ran into

Multi-model orchestration timing was the biggest challenge. Routing a single user message through three sequential model calls (Hermes triage, then conditionally K2 reasoning, then Claude response) while keeping the total response time under a few seconds required careful prompt engineering to keep each call focused and concise. We optimized by making the Hermes classification call return minimal JSON rather than natural language, and by only invoking K2 when the triage indicated a rights or compensation query.

Designing for accessibility without being patronizing was harder than expected. Our target users are capable adults who happen to have limited digital literacy or English proficiency. Every design decision had to respect their intelligence while removing unnecessary complexity. We went through several iterations on the rebooking comparison cards before finding a layout that was information-dense but not overwhelming.

Getting the DOT compensation reasoning right required significant prompt engineering for K2. Airline compensation rules have edge cases and exceptions that a generic prompt misses entirely. We had to enumerate specific scenarios (voluntary vs involuntary bumping, controllable vs uncontrollable delays, domestic vs international) and structure the reasoning to handle each branch correctly.

Making the demo flow feel real in a hackathon context was its own engineering challenge. We built a complete automated walkthrough that simulates a flight cancellation in real-time, showing the status change animations, toast notifications, alternative search, and compensation calculation in a smooth 55-second sequence that captures the full user experience.

Accomplishments that we're proud of

Three-model orchestration through a single gateway — Hermes classifies, K2 reasons, Claude communicates, all through one Lava API endpoint. This is not three separate API integrations duct-taped together. It is a genuine pipeline where each model's output feeds the next.

Transparent legal reasoning — Users can see K2's full step-by-step reasoning chain for their compensation calculation. This is not a black box that says "you're owed $650." It shows you exactly which DOT rules apply and how the number was calculated, giving workers the confidence and evidence to actually claim what they are owed.

Voice-first multilingual design — The entire app is usable without touching the keyboard. A Spanish-speaking construction worker can open ReRoute, tap the microphone, describe their situation in Spanish, and get a full response with rebooking options and legal rights information in Spanish.

Cross-modal transportation search — Airlines only show you their own flights. ReRoute shows you buses, trains, rideshares, and combination routes. For a worker on a tight budget, the $65 Amtrak option that the airline never mentions could be the difference between making their shift and losing a day's pay.

What we learned

Blue collar workers are systematically underserved by travel technology. Every major travel app and airline portal is designed for frequent flyers with credit cards, loyalty programs, and flexible schedules. Building for workers who fly twice a year on a tight budget with inflexible work commitments required rethinking almost every assumption about how travel tools should work.

Multi-model orchestration is more powerful than single-model prompting. Splitting triage, reasoning, and communication across specialized models produced dramatically better results than trying to get one model to do everything. The Hermes classification layer alone reduced irrelevant K2 invocations by routing simple conversational messages directly to Claude.

Showing your work builds trust. When we added the K2 reasoning chain display to the compensation calculator, it transformed the feature from a novelty into something a worker could actually use at an airline counter. Transparency is a feature, not a debugging tool.

Voice is not a nice-to-have for accessibility. It is the interface. For our target users, voice input is not an alternative to typing. It is the primary interaction mode. Designing voice-first rather than adding voice as an afterthought changed the entire UX.

What's next for ReRoute

Real flight data integration — Connecting to FlightAware or AeroAPI for real-time flight status monitoring instead of mock data, enabling truly proactive disruption detection.

Automated rebooking execution — Moving beyond comparison to actually booking alternatives on behalf of the user, integrating with bus, train, and rideshare booking APIs.

SMS and WhatsApp interface — Many blue collar workers rely on text messaging more than apps. We want to make ReRoute accessible via SMS and WhatsApp so workers do not need to install anything.

Employer integration — Partnering with staffing agencies and employers so ReRoute can automatically notify supervisors about travel disruptions and coordinate schedule adjustments.

Compensation claim automation — Filing DOT compensation claims on behalf of users, handling the paperwork and follow-up that most workers do not have time for.

Expansion beyond flights — Applying the same advocacy model to bus cancellations, train delays, and other transportation disruptions that affect working people.

Built With

• claude
• framer-motion
• hermes
• hex
• k2-think-v2
• lava
• next.js
• recharts
• shadcn-ui
• tailwind-css
• typescript
• vercel
• web-speech-api