1. Duplicate Activities Across Days

The Problem:

When generating multi-day itineraries, Days 1, 2, and 3 were showing identical activities. Users would get the Eiffel Tower repeated three times instead of a diverse travel experience.

Root Cause:

Each day was making separate API requests without tracking what had been used
No global state to prevent activity repetition
Distribution algorithm was naive and didn't share context across days

How I Fixed It:

Caching for 24 hours responses and filtering

2. API Rate Limiting Issue

The Problem:

OpenTripMap API was returning 429 (Too Many Requests) errors during itinerary generation, causing the entire flow to fail.

Root Cause:

Making rapid-fire requests (one per category per day)
No retry mechanism
No exponential backoff

How I Fixed It:

Exponential backoff and first category analysis and identification through ML

Impact:

Rate limit errors dropped from ~40% to <1% of requests.

3. Google Places API Migration Complexity

The Problem:

OpenTripMap had limited data quality and outdated information. Migrating to Google Places required mapping user intents to 100+ place types.

The Challenge:

Google Places has 122 different place types (museum, art_gallery, tourist_attraction, etc.)
User queries like "show me art in Paris" needed to map to multiple types
No built-in category detection

How I Fixed It:

Created google-places-types.ts with comprehensive type mappings
Used GPT-4o-mini for intelligent intent-to-type detection
Fallback to default types if LLM fails

Impact:

Search relevance improved by 85% with real-time, accurate place data.

4. Socket.io Real-Time Sync Issues 📡

The Problem:

Chat messages were out of order or duplicated due to race conditions between Socket.io and REST API calls.
Frontend was making both REST POST and listening to Socket.io events
No message deduplication
Multiple socket connections per user

How I Fixed It:

Frontend: Deduplicate messages by ID
single socket per connection

Impact:

Message sync issues reduced from ~15% to 0%.

5. LangGraph State Management

The Problem:

State wasn't updating correctly between LangGraph nodes. The searchResults field was being overwritten by subsequent nodes.

Root Cause:

Misunderstanding of LangGraph's reducer functions
Using reducer: (left, right) => right for arrays
Not returning partial state updates correctly

How I Fixed It:

Used concatenate and partial state updates were fixed

Impact:

State flow became 100% predictable with no data loss.

6. Image Fallback Problem

The Problem:

Activity cards were showing random Unsplash stock photos instead of real place images from Google Places API.

Root Cause:

Multiple fallback layers (Google → Unsplash → placeholder)
Old code still generating Unsplash URLs
Frontend had aggressive fallback logic

How I Fixed It:

Google Places plus OpenTripMap fallback

Impact:

100% authentic Google Places photos displayed. No more confusing stock images.

7. Flight API Integration Complexity

The Problem:

Amadeus Flight API required IATA codes (JFK, CDG), but users enter city names. Also, their access token expires every 30 minutes.

How I Fixed It:

OpenAI LLM basedweb-search fallback for lookuo

Impact:

IATA lookup success rate: 98%.
Token refresh errors: 0. API cost reduced by 80% with caching.