Plexis - Trips you can trust, for Alexa+

An Alexa+ travel agent that only offers plans it can prove: every trip is solved, independently verified and signed, and impossible requests get the exact clash plus a fix.

Comment

Inspiration

Ask a voice assistant to plan a family trip with a budget, a wheelchair user and a vegetarian, and you get a confident itinerary. Then you find the hotel has steps at the entrance, the restaurant is closed that night, and the total is over budget. When an assistant spends real money, confident isn't enough. It has to be right, and it has to say so when what you asked for can't be done.

What it does

Plexis is an Alexa+ add-on that plans complete trips (flights, stays, meals and activities) and only offers plans it can prove. • A real conversation, by voice. “Book a trip for me” starts a dialogue that asks one thing at a time (where, when, who, from where, budget, room needs), with three spoken suggestions. “The first one”, “3” and “yes” are understood in context. • Asks about what people forget. About 30 researched preferences: pool, spa, gym, breakfast included, all-inclusive, free parking, kitchen, view, quiet room, crib, AC or no AC, shared or private room. Each can be wanted or explicitly unwanted. • Proof, not promises. A constraint solver finds the best plan, an independent checker re-verifies every rule, and a second solver cross-checks it. The result is signed into a tamper-evident receipt. • Impossible requests get a fix, not a made-up plan. It names the smallest set of clashing rules, and every suggested fix is re-solved before it's offered. • Why not that hotel? It answers by showing exactly which of your rules that choice would break. • A family member that learns. Household memory with consent, phrases you teach it (“when I say beach vibes, I mean beachfront with a pool”), corrections it logs, and habits it notices. • It keeps watching. If a booked flight is cancelled, it re-plans around your hotel and tells you. It also suggests drier dates, a packing and outfit brief, and a “don't miss” place.

How we built it

alexa+ add-on: a self-hosted MCP server (Streamable HTTP, spec 2025-11-25) following the Alexa+ Toolkit conventions: structured output with an output schema, an MCP Apps card for Echo Show, confirmation by elicitation before payment, duplicate-booking prevention, and two-tier OAuth (client credentials and PKCE). • LangGraph runs every conversation turn as a state graph. The planning graph runs seven partner and research lookups in parallel, then solving, then verification and the cross-check side by side. Partner data is prefetched while the customer is still answering. • OR-Tools CP-SAT solves the plan with one assumption per rule, so impossible requests yield a minimal conflict. Z3, or an arithmetic cost floor, independently confirms results. Receipts are signed with Ed25519 into a hash-chained ledger. • Amazon Bedrock with Nova, through the Strands Agents SDK, handles phrases the deterministic parser doesn't understand. Each proposal must quote the customer's own words. • Two simulated Alexa+ hosts, both real MCP clients: a voice-only device (offline Whisper and system voices) and a web simulator showing the card, protocol trace, latency waterfall, in-browser ledger verification and a chaos lab. • Live weather from Open-Meteo; destination highlights from curated landmarks plus Wikivoyage.

Challenges we ran into

• The Alexa+ MCP Toolkit is partner-only, so we built our own simulated hosts. • The 401 rule in the Alexa+ docs conflicts with the MCP spec. • The Python and TypeScript MCP SDKs are on different major versions. • Z3 timed out on budget proofs, so we added an arithmetic certificate. • Windows speech needs its own thread, and an always-open mic answered the TV, so we added press-to-talk. • A self-audit found that “proven” was only as good as the rules: seats left, rental-car return and time zones were missing. All three checkers now enforce them.

Accomplishments that we're proud of

Every plan it offers comes with a proof, every impossible request comes with a minimal explanation and a verified fix, and the whole thing works hands-off by voice.

What we learned

Trust comes from checking, not from fluent language. Keeping the model at the edges, quoting the customer's words and verified against the plan, makes an agent that is both helpful and safe to let spend money.

What's next for Plexis - Trips you can trust, for Alexa+

Deploy on Amazon Bedrock AgentCore Runtime, run Bedrock live, connect real flight and hotel APIs through AgentCore Gateway, and bring the same approach to group dinners, events and errands.

Built With

Share this project:

Updates

Submission history