Skip to main contentSkip to footer
    Back to Case Studies
    Beyond Points AI logo
    AI AgentsTravel & LoyaltyWeb Platform

    Beyond Points AI

    From a points dashboard to an AI travel strategist that finds the smartest way to book a trip, then books it — with a human confirming every irreversible step.

    Project Preview

    Live website preview - beyondpoints.ai

    13
    Tools the AI chat can call
    71
    Supabase edge functions
    3
    MCP servers (browser, loyalty, travel)
    0%
    Error rate, 30-search live run

    Conversational travel strategist

    Streaming chat on Gemini 2.5 Flash with 13 callable tools: award search, cash fares, hotel calendars, transfer math, itinerary optimization.

    Award + cash flight search

    Cross-checks seats.aero award availability against Travelpayouts and Google Flights cash fares, party-aware and multi-city per leg.

    Hotel award intelligence

    Per-date points and cash on the same record for chain programs, so the optimizer can score cents-per-point honestly.

    Automated point transfers

    A browser agent completes card-to-partner transfers end to end, behind an explicit confirmation gate.

    Confirmation gates everywhere

    Transfers, bookings, payments and cancellations never execute without a user checkpoint.

    Bank-grade account security

    TOTP MFA, nine OAuth providers, AES-encrypted tokens, 15-minute idle and 24-hour hard session limits, RLS on every table.

    The founder's problem

    Gregory Zaks, founder of Beyond Points AI, was sitting on a good idea and a half-built product. Travelers hold hundreds of billions of dollars of unused points every year, and the people who do redeem them well are experts in transfer-partner matrices, award calendars and sweet spots that change monthly. Everyone else either books cash and lets points rot, or burns them at a fraction of their value.

    Beyond Points already had a Supabase-backed platform that could show a user their balances. What it could not do was the hard part: reason across every account a traveler holds, compare award inventory against cash fares in real time, and then actually move points and complete a booking. That work was still manual, across a dozen airline and hotel sites, and it was the whole value proposition.

    Why Beyond Points came to Frenchy Digital

    • An agent layer, not a chatbot: The brief was a fully autonomous booking platform that transfers points and completes bookings, with user checkpoints, on top of the existing Supabase backend.
    • Judgment as the moat: In the founder's words, Beyond Points is not a search engine. It has to explain why one option beats another, in cents per point, and be right.
    • Speed with guardrails: The product touches real money and real loyalty balances. Automation had to be fast and still impossible to run without a human's say-so.

    How AI is used

    Two AI layers with different jobs

    The conversational layer runs on Google Gemini 2.5 Flash with server-sent streaming. It exposes thirteen tools to the model, from award search and hotel calendars to transfer math and itinerary optimization, so the chat never guesses at a fare it could look up. Dedicated edge functions handle points optimization and itinerary optimization as separate, testable calls.

    The autonomous layer is a Claude-based multi-agent system speaking the Model Context Protocol (MCP). An orchestrator delegates to loyalty, transfer, browser, flight-booking and hotel-booking agents, plus a code-reviewer agent and a set of fifteen executable skills. The cost model pairs a Sonnet-class orchestrator with Haiku-class sub-agents so that long-running work stays cheap.

    Three MCP servers

    browser-mcp

    Headless Chromium in Docker on a VPS behind Caddy with a residential proxy. Scrapes balances and award availability, executes transfers and bookings on airline and hotel sites.

    loyalty-mcp

    Program rules, transfer partners and ratios, direct Amex, Chase and Capital One rewards integrations, OAuth or credential-based balance retrieval.

    travel-mcp

    seats.aero award inventory with seat counts, Travelpayouts cash fares, Google Flights as a fallback, and AwardTool's hotel API for per-date points and cash.

    The journey optimizer

    Search results feed a gateway planner that assembles hybrid points-plus-cash journeys, considers nearby airports and flexible dates, and scores each candidate in cents per point. The ranking is deterministic: eight identical concurrent searches return byte-identical results, with session de-duplication, cancellation via AbortSignal and stale-run guards so a slow provider can never overwrite a newer answer.

    How we built it

    The original proposal sketched a microservice fleet on Kubernetes with a separate planning model. We rejected it during discovery. The founder needed a product he could operate, not a platform team. What shipped is a React and TypeScript single-page app on Cloudflare Pages, a single Supabase project carrying Postgres, Auth, Realtime and 71 Deno edge functions across roughly 70 migrations, and one VPS running the browser agent. Every table has row-level security.

    AI travel chat with streaming responses and tool calls
    Award and cash flight search with multi-source cross-checking and party-aware pricing
    Hotel search with per-date award and cash pricing
    Browser-agent flight and hotel booking behind confirmation gates
    Points portfolio and wallet with Realtime updates, plus a transfer calculator
    Automated point transfers end to end via the browser-mcp bridge
    Point Map: a visual galaxy of every program a user holds
    Deals hub with alerts, itinerary builder with AI optimizer, six-stage /plan funnel
    Professional partner program, enterprise inquiries, full admin and super-admin
    Free, Pro and Enterprise tiers, admin-granted, with usage limits on AI inquiries

    Timeline

    1. October 2025: First commit. Core platform, auth and points portfolio.
    2. February 2026: Agent-layer engagement scoped: autonomous search, transfers and booking with user checkpoints, on an eight-week launch plan.
    3. June 2026: Infrastructure cutover to Cloudflare Pages, a fresh Supabase project and the browser-agent VPS. Architecture docs consolidated.
    4. August 2026: Sprints 3 to 5: optimizer hardening, live provider validation, timing-claim remediation, vendor de-risking, founder beta readiness.

    Problems worth talking about

    Silent adapter drift

    An evidence adapter written to the chat's data shape was defaulting every live leg to economy with a blank carrier. Nothing crashed, the results were just quietly wrong. We fixed it and locked it with a unit test that fails if the shapes drift again.

    Claims that could not be proven

    An audit found 112 timing assertions in the product, 83 of them unlabeled. We built a four-category claim taxonomy and swept every one so the UI only says what the logs can back up.

    An 11-day email outage nobody saw

    Two stacked SMTP misconfigurations stopped signup and reset emails for eleven days, masked by Google OAuth still working. The incident is written up in the repo, with the monitoring that now catches it.

    Fewer vendors, better data

    Duffel, Bright Data, Amadeus, Plaid and Stripe were all removed. Hotel and cash sources were re-sourced with a capability matrix that documents what each provider can and cannot return.

    Where it stands

    Measured, not marketed

    In the pre-beta live run of thirty real searches, the platform averaged 4.7 orchestrator calls and about nine HTTP calls per search, returned ranked journeys for 29 of 30 queries with a 3.3% empty-result rate and a 0% error rate, and measured cents-per-point on 25 of the 29. Median wall time was 20 seconds. Those are the numbers we publish because they are the ones we can reproduce.

    Beyond Points is in founder beta. We deliberately do not quote user counts, bookings or dollars saved here, because the product has not opened to the public yet. When it does, this page will be updated with real figures.

    Not a search engine. Our moat is judgment.

    — Beyond Points AI product principle

    What This Project Does Well

    Live multi-source search: 29 of 30 real searches surfaced ranked journeys, with a 0% error rate
    Eight identical concurrent searches produce byte-identical rankings (deterministic optimizer)
    Point transfers automated end to end through the browser agent bridge
    Every irreversible action (transfer, book, pay, cancel) sits behind a confirmation gate
    Vendor risk cut: Duffel, Bright Data, Amadeus, Plaid and Stripe removed and re-sourced with a documented capability matrix
    Cost model on Claude Sonnet orchestrator + Haiku sub-agents keeps autonomous work affordable

    Technologies Used

    React 18, TypeScript, Vite, Tailwind CSS, shadcn/ui, TanStack Query, Supabase (Postgres, Auth, Realtime, 71 Deno edge functions, ~70 migrations, RLS), Google Gemini 2.5 Flash, Claude API + MCP multi-agent system, Puppeteer browser agent in Docker on a VPS behind Caddy, seats.aero, Travelpayouts, SearchAPI, AwardTool Hotel API, Amex / Chase / Capital One rewards APIs, Cloudflare Pages, Resend

    Ready to Build Something Similar?

    Let's discuss how we can help transform your business with a custom digital solution.

    Open beyondpoints.ai📅 Schedule Consultation