I built a travel app with Abstract Factory + affiliate APIs — here's what actually worked

Sivaraman Ramaswami — Thu, 12 Nov 2026 14:54:41 +0000

I've been building packedNbooked (https://packednbooked.com) — a trip planner that starts with your budget instead of a destination. You enter $2,000, JFK, 7 days → it fans out to flight/hotel/car/activity APIs in parallel and returns ranked complete trip packages.

This post is about the architecture decisions that held up and the ones that didn't.

The pattern: Abstract Factory

Every travel category (flights, hotels, cars, attractions) gets its own factory. Each factory spawns provider agents. The orchestrator never touches provider logic directly.

class FlightProviderFactory(BaseFactory):
def create_agent(self) -> BaseAgent:
return TravelpayoutsAgent()

class HotelProviderFactory(BaseFactory):
def create_agent(self) -> BaseAgent:
return BookingAgent()

This paid off when I swapped hotel providers mid-project (Hotellook → Booking.com via CJ affiliate). Zero changes to the orchestrator — I just swapped the agent behind the factory. Same when Ticketmaster went live: new factory, new agent, plug it in.

The fan-out itself runs async — all four factories fire in parallel, results get budget-pruned, then an LLM (Groq + gpt-oss-120b) ranks the combos by value.

The hotel API nightmare

This is the part nobody writes about.

My original plan was real-time hotel pricing. Here's what I found:

Amadeus — decommissioned their free tier in July 2025
Travelpayouts hotel endpoints — all 404
Booking.com, Agoda, Expedia, HotelBeds — all require commercial contracts
Booking.com via CJ affiliate — Demand API is restricted to "strategic partners," no application path

I spent three weeks on this before accepting the ceiling: mock prices + live Booking.com deep-links is the permanent setup. The links go to Booking.com search results where users see real availability. It's honest — I label prices as estimates and the Show HN caveat was upfront about it.

The lesson: travel pricing APIs are a moat. If you're not paying enterprise rates, you're not getting real-time hotel data. Build around it, not through it.

The affiliate setup (actual numbers)

Revenue model is affiliate commissions — no ads, no paywall:

Partner	Integration	Notes
Travelpayouts/Aviasales	Live prices via Data API	marker 326901
Booking.com via CJ	Deep-links, mock prices	publisher 101876939
Viator	Live affiliate links	ID P00319033
DiscoverCars	Affiliate link	~$20/booking, 365-day cookie

Booking.com's CJ approval took about a week. Viator was instant. DiscoverCars was the fastest approval and best cookie window.

Realistic projection: $0–20/mo now, $200–430/mo at 20k monthly visitors. It's a long game — the affiliate revenue only matters if the SEO compounds.

What's actually live

FastAPI backend on Railway ($5/mo), React/Vite frontend on Render (free static)
Flights: real cached prices from Travelpayouts /v2/prices/latest
Hotels: estimated prices, live Booking.com deep-links
Attractions: Viator (live) + Ticketmaster Discovery API (real events by city/date)
AI ranking: falls back to cost-sort if Groq key is missing
115 tests passing

What I'd do differently

Start with the affiliate approvals. I wired affiliate links at the end and retroactively got approvals. Should have done it week one — some take a week+, and you don't want to build around a partner that rejects you.

Don't fight the hotel API wall. I lost three weeks trying to find real-time hotel pricing. The correct answer was always "mock prices + live booking link."

Abstract Factory was the right call. The pattern felt like over-engineering for a solo project at first. It wasn't — I swapped providers three times without touching orchestration logic.

App is live at https://packednbooked.com. Code isn't public yet but happy to answer questions about the architecture or affiliate setup.

What travel pricing APIs have you found that actually work at the free/indie tier?

DEV Community: Sivaraman Ramaswami

I built a travel app with Abstract Factory + affiliate APIs — here's what actually worked