SimplifAI Travel

Problem & Motivation

Have you ever planned a trip for yourself or your family and friends? 

If yes, how much time did it take you to plan and create an itinerary and list of to-dos? If not, you’re the lucky ones.

Did you know that average trip planning takes at least 16-20 hours. 

Itinerary planning can be time consuming and stressful, especially if it is your first time going to a travel destination. The endless web searching to learn about the destination, countless hours of reading restaurant reviews, and trying to map out how to move from attractions to attractions can be a tough task to navigate through. 

As travel enthusiasts ourselves, we are also in search of the perfect travel placing tool that can help us get through the tedious process of itinerary planning. Wouldn’t it be nice if through a couple clicks of the buttons and inputting our interest, someone will auto-populate a well-planned itinerary with attractions, restaurants and best routes to take? Better yet, it will consider the weather at the destination and adjust the itinerary accordingly. Didn’t we all encounter situations where we planned for an outdoor attraction and weather ruined it all. 

Here, we want to introduce you to SimplifAI Travel, your personal AI-powered travel agent that will help you plan your next trip.

Our Solution

SimplifAI Travel is an AI travel tool based on a multi-agent framework that allows users to simply input their preference and generate an itinerary based on the date, interest and a few considerations (like weather or traffic) they picked for their destination.

We hope through SimplifAI Travel, users can minimize the time used to plan their next big trip from hours to minutes. SimplifAI Travel also has a feature that allows users to explore key attractions for the selected destination. This is always helpful for anyone who is traveling to the destination for the first time.

Product Features

SimplifAI Travel, is designed to deliver intelligent, real-time itinerary planning tailored to individual user preferences. The application generates full-day travel schedules from 9:00 AM to 9:00 PM, ensuring a balanced mix of activities and downtime. It incorporates constraints such as a maximum of three hours per activity and includes lunch and dinner plans at well-reviewed restaurants located within a 10-mile radius of the planned attractions. These dining recommendations are filtered to match operational hours and user preferences, and the system avoids repeating venues across the trip.

To further enhance the travel experience, the system integrates weather forecasts to influence activity timing and suggest relevant packing items. In the case of inclement weather, it proposes suitable indoor alternatives. Each itinerary also includes driving directions, estimated travel times, and daily themes to guide the user experience.

The route optimization component considers real-time traffic conditions to minimize distance, time, and disruptions, while avoiding inefficient routing patterns such as backtracking. These features work collectively to create a personalized, seamless, and efficient trip planning experience.

The table below outlines the categories and features of this product:

Table 1: Product Features 

Data Science Approach

Exploratory Data Analysis (EDA)

This project leverages real-time data from two primary APIs: Google Places API for point-of-interest and dining information, and Open-Meteo API for hourly weather forecasts. The collected data is structured and includes essential attributes such as location, business status, user ratings, editorial summaries, opening hours, cuisine types, and weather conditions. These features form the foundation for constructing personalized and context-aware travel itineraries.

Google Places API provides access to a vast and frequently updated dataset, covering over 109,000 cities across the United States. This makes it suitable for large-scale applications requiring geographical diversity and high data reliability (Google Developers, 2023). The Open-Meteo API offers free, high-resolution weather forecasts and supports integration into automated planning systems, enabling adaptive itinerary generation based on hourly and daily conditions (Open-Meteo, 2023).

During the exploratory data analysis (EDA) phase, efforts were made to ensure data quality by filtering irrelevant or incomplete entries and identifying fields critical for LLM-driven decisions. The most influential fields included rating scores, descriptive summaries, child-friendliness indicators, and weather descriptors. These features were selected for their impact on content generation, user relevance, and itinerary feasibility.

To align system output with user interests, a mapping schema was developed that connects general preferences (e.g., “Nature,” “Shopping”) to specific API search categories (e.g., “park,” “zoo,” or “shopping_mall”). This semantic mapping optimizes query precision and enhances recommendation accuracy.

Together, the use of high-quality, real-time API data and structured semantic mapping supports the core objectives of this product: to build a responsive, personalized, and intelligent AI-driven trip planning system.

Table 2: Dataset Metadata

Table 3: User's Interest to Search-Types Mapping

Data Pipeline

To ensure high-quality and context-aware itinerary generation, this system employs structured data pipelines tailored to three core travel planning components: attractions, dining, and weather. All data retrieved from external APIs is structured in JSON format before being passed to the Large Language Model (LLM). This method reduces prompt length, minimizes hallucinations, and improves both relevance and accuracy of model responses.

Attractions Data Pipeline

The attractions pipeline is responsible for selecting locations that best align with the user’s stated interests. The process begins with the user's destination and preferences, which are mapped to specific Google Places API search types (e.g., “museum,” “zoo,” “tourist_attraction”). The pipeline then calls the API, and the returned results are filtered using LLM prompts to surface the most relevant options. This approach supports adaptive query interpretation and ensures personalized outputs (Google Developers, 2023).

Nearby Restaurant Data Pipeline

For each selected attraction, the system finds suitable dining options within a 10-mile radius. The API search includes categories such as “restaurant,” “diner,” “food court,” and “steakhouse.” The results are again filtered through the LLM using prompts and decision heuristics based on ratings, open hours, and proximity. This ensures that users receive dining recommendations that are both convenient and contextually appropriate.

Weather Forecasts Data Pipeline

Weather forecasts are critical for trip planning, especially for outdoor activities. The pipeline retrieves hourly forecast data from the Open-Meteo API for the specified travel dates and locations. The raw data includes temperature, time, and WMO (World Meteorological Organization) weather codes. These codes are converted into descriptive labels (e.g., “Sunny,” “Cloudy”) to aid interpretability and to support itinerary adjustments in response to unfavorable conditions (Open-Meteo, 2023).

This multi-pipeline structure ensures that the LLM receives clean, well-defined, and timely inputs for generating accurate, reliable, and user-specific travel plans.

ML Pipeline & Techniques

This project adopts a multi-agent AI system to decompose the problem of trip planning into smaller, manageable tasks. Each task is assigned to a specialized AI agent, forming a modular and collaborative machine learning pipeline. This design promotes specialization, parallel execution, and improved interpretability.

Problem Decomposition and Agent Assignment

The overall task of generating a personalized travel itinerary is divided into four sequential steps: (1) selecting attractions based on user preferences, (2) finding restaurants near those attractions, (3) retrieving weather forecasts for travel dates, and (4) building and optimizing the final itinerary. Each step is handled by a designated agent using the CrewAI agent framework.

Agent-Based Task Execution

• Task 1: Attraction Selection
   The “Local Tour Guide” agent queries Google Places API using mapped search terms derived from user interests. The agent filters results using an LLM with guardrails to ensure relevance and accuracy.
• Task 2: Restaurant Recommendation
   The “Restaurant Scout” agent finds nearby restaurants within a 10-mile radius of each attraction. It uses API search types such as “restaurant,” “diner,” or “food court,” and applies similar filtering via the LLM.
• Task 3: Weather Retrieval
   The “Meteorologist” agent accesses the Open-Meteo API to obtain hourly weather forecasts. Weather codes are converted to human-readable descriptions to support planning decisions and enhance the itinerary’s contextual awareness.
• Task 4: Itinerary Optimization
   The “Trip Planner Specialist” agent assembles all gathered data—attractions, restaurants, and weather—into a full-day itinerary. The plan includes details such as activity timing, reviews, dining options, and packing suggestions. Guardrails are again applied to maintain consistency and quality.

Output Integration

The final output consists of a structured daily itinerary that includes activity information, weather conditions, meal recommendations, driving directions, and personalized packing tips. This agentic pipeline facilitates dynamic and user-centric travel planning by integrating real-time APIs and structured reasoning.

Diagram 1: Problem Decomposition & Multi-AI Agent System

Step-by-Step Explanations:

1. User preferences are sent to the Local Tour Guide agent to find matching attractions.
2. The agent uses the Google Places API and LLM to filter results, with guardrails ensuring quality.
3. Selected attractions go to the Restaurant Scout agent to find nearby dining options for each attraction.
4. Google Places API is used again to suggest restaurants. LLM + guardrails ensure relevance.
5. With attractions and restaurants ready, we move to Task 3, assigned to the Meteorologist agent, which pulls weather data for the travel dates.
6. The agent pulls hourly forecasts from Open-Meteo and formats them for easy use.
7. All inputs, attractions, restaurants, and weather, are passed to the Trip Planner Specialist agent.
8. This agent builds the full itinerary, checking flow, timing, and balance.
9. The plan is validated using LLM with guardrails to catch errors and ensure consistency.
10. The final daily itinerary is generated, including activities, meals, weather, driving info, and packing tips.

Modeling Approach & Selection

To support real-time, personalized trip planning, this project adopts an agentic AI architecture instead of traditional custom-trained models. Custom models often struggle to access live data (e.g., weather, traffic, business hours) and require significant data volume, computational resources, and frequent retraining to stay current. Additionally, such models are less flexible in adapting to dynamic user preferences and have slower iteration cycles compared to lightweight, API-based approaches.

In contrast, the agentic AI framework enables modular, real-time task execution by leveraging pre-trained language models, external APIs, and structured orchestration. Each agent performs a specific task—such as retrieving attractions, filtering restaurants, or analyzing weather forecasts—improving system efficiency and clarity.

Moreover, the system uses a multi-agent design rather than a single-agent setup. Multi-agent systems allow for task specialization, which enhances performance, scalability, and fault tolerance. Agents operate in parallel, accelerating response time, and better reflecting how humans typically break down complex planning processes into smaller steps (Yao et al., 2023). This approach improves flexibility and makes the system easier to debug, test, and extend.

Model Insights: Features, Strengths & Limitations

The performance of the Real-Time AI Trip Planner is driven by several key features. The most significant among these are user preferences, location data, restaurant availability, and weather forecasts. User preferences play a critical role by enabling personalized itinerary generation tailored to individual interests and priorities. Location proximity affects routing efficiency, minimizing travel time between activities. Meanwhile, weather data ensures time-appropriate and condition-sensitive activity planning, improving overall experience quality.

The model performs optimally in urban areas with dense clusters of attractions and ample dining options. These environments, when paired with consistent and accurate API-sourced data, support high personalization accuracy and effective real-time adaptation.

However, the model exhibits limitations in low-density or rural regions where API data may be sparse or outdated. It also faces challenges when user preferences conflict—such as combining nightlife with family-friendly activities—or when attractions are widely dispersed, potentially impacting itinerary coherence and travel efficiency.

Key Technical Innovations: From System Design to User Experience

This project demonstrates how thoughtful system design can significantly enhance user experience in AI-powered travel planning. Three core technical innovations underpin the application’s performance and responsiveness.

First, the system reduces hallucination in large language model (LLM) outputs by preloading structured inputs and incorporating validation guardrails, improving accuracy and trustworthiness in itinerary generation. Second, performance is boosted through asynchronous task execution and efficient data handling, allowing multiple agents to work in parallel and reduce latency. Third, by replacing static datasets with real-time API integrations, the system adapts dynamically to current conditions such as weather, traffic, and business status, offering personalized and timely recommendations (Google Places API, 2024; Open-Meteo, 2024).

Unlike traditional itinerary tools, this solution adopts an agentic AI architecture, where task-specific agents manage modular components like attraction search, restaurant selection, and itinerary assembly. This architecture, built using frameworks like CrewAI, enhances adaptability, scalability, and realism—mimicking how humans plan trips step by step.

By combining technical robustness with user-centric design, the system delivers an intelligent, reliable, and engaging trip planning experience.

Overall Architecture

The deployment architecture of the SimplifAI Travel application is designed for scalability, modularity, and real-time responsiveness. It is hosted on Google Cloud and containerized using Docker, orchestrated via Google Kubernetes Engine (GKE) to ensure efficient scaling and fault tolerance.

The architecture consists of three primary components:

1. Frontend Layer: The user interface is built with Streamlit, allowing users to input trip preferences and visualize itinerary results. The frontend communicates securely with the backend over HTTPS.
2. Backend Layer: A FastAPI server handles all trip planning requests. This layer integrates the CrewAI multi-agent framework, which orchestrates task-specific agents responsible for selecting attractions, finding restaurants, retrieving weather data, and generating the final itinerary.
3. External Services: The system utilizes external APIs for live data:
   • Google Places API for real-time information on attractions and restaurants
   • Open-Meteo API for hourly weather forecasts
   • OpenAI’s GPT model for natural language processing and itinerary generation
4. Evaluation and Observability: Tools such as Langtrace and Ragas are used to monitor, evaluate, and trace the performance and reasoning paths of the agents and LLMs. This ensures reliability, reduces hallucination, and supports continuous model improvement.

This modular deployment strategy allows for independent scaling of the frontend and backend, real-time API access, and robust system monitoring—ultimately leading to a smoother and more intelligent user experience in travel planning.

Diagram 2: Application Deployment Architecture

Model Evaluation

To ensure the quality, reliability, and utility of responses generated by AI agents within a travel assistant setting, we developed a structured evaluation pipeline using the Ragas framework. This approach allows us to assess the performance of agent-based systems in realistic user scenarios, benchmarking them across multiple qualitative metrics.

Our evaluation is anchored in real user interactions collected from the SimplifAI Travel platform—an AI-powered travel planner designed to deliver personalized itineraries and recommendations. We sampled 30 user inputs that varied across travel planning, local activity discovery, and weather-related queries. These inputs provide the base context for agent responses.

The context module processes each user input to generate a richer, situational prompt, which is then handed to a crew of agents coordinated by Gemini 1.5 Pro. This system of agents is tasked with not only generating dynamic, context-aware responses but also producing golden answers—reference outputs considered ideal for benchmarking.

Diagram 3: RAGAS

To assess the effectiveness of different agent configurations, we ran comparative evaluations using agent stacks built on ChatGPT-4o-mini and Gemini 2.0-Flash, operating at multiple temperature levels (0, 0.5, 0.9). This allowed us to analyze how deterministic versus more exploratory agent behaviors impacted the final outputs. All agent responses were then passed through the Ragas framework.

Ragas evaluated the responses using three key metrics:

• Faithfulness: Does the agent’s output adhere to the facts and context provided?
• Response Relevancy: Is the response aligned with the original user intent?
• Response Accuracy: Are the recommendations or statements factually correct?

To benchmark results, we used ChatGPT-4 as a reference model—representing a high-quality control agent setup for comparison.

This modular evaluation approach enables us to isolate the performance impact of prompt engineering, agent orchestration, temperature tuning, and underlying model selection. It ensures a robust and fair evaluation of agents deployed in a real-world application setting.

In conclusion, this agent-focused framework powered by Ragas provides a scalable and consistent methodology to assess and iterate on intelligent agent performance, helping us improve both the backend logic and end-user experience in domain-specific AI applications like travel planning.

Diagram 4: RAGAS Results

Gemini 1.5 Pro demonstrated the highest faithfulness score at 0.98, indicating a near-perfect alignment between generated content and the original context. However, its accuracy score could not be computed (marked N/A) due to formatting inconsistencies in golden answers.
ChatGPT-4o-mini also performed strongly on faithfulness (0.97) and relevancy (0.76), but had the lowest accuracy score (0.47), suggesting that while the outputs were on-topic and fact-aligned, they often lacked factual precision.
Gemini 2.0-Flash showed balanced but slightly lower performance across all two dimensions: faithfulness (0.96), relevancy (0.71), while accuracy (0.69) was higher.

From these results, we observe that faithfulness remains consistently high across all agents, suggesting the context-prompting strategy and agent structure are well-optimized to stay grounded.

Key Learnings & Impact

This evaluation highlights how agent-based architectures can achieve strong contextual alignment, but also underscores the need for improved factual grounding—especially when planning travel, where precision matters (e.g., locations, weather patterns, or event times). By continuously refining agent tools and prompts, and leveraging insights from frameworks like RAGAS, we can iteratively improve the end-to-end reliability and utility of AI-driven assistants in real-world applications

Further, the use of the Agentic AI can create personalized realtime itineraries that source information from a variety of sources and drastically reduce the cognitive load on families and individuals who want to plan trips. 

Roadmap

We hope in the future iterations of SimplifAI Travel to include the following features:

• User feedback loops to refine itinerary quality over time.

• Chatbot interface to allow users to refine the itinerary to align with their personal needs.

• Expand capabilities to allowing processing of visas, buying of services (airline tickets, ground transportation, attraction tickets)

• Expand support for multi-user trip planning and collaborative scheduling.
   

Generalization

The agentic AI approach is flexible and can be adapted to other domains like event planning, logistics, or scheduling—anywhere tasks can be broken down and handled by specialized agents. We planned our product to be comprehensive of all travelers' need.

Acknowledgement

We would like to thank our instructors Danielle Cummings and Todd Holloway for their feedback, guidance and invaluable insights throughout our journey.

References

• Google Developers. (2023). Places API overview
• Open-Meteo. (2023). Weather API
• Yao, S., Zhao, J., Yu, D., et al. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. NeurIPS.
• https://press.priceline.com/new-priceline-research-finds-average-traveler-spends-two-full-work-days-to-plan-and-book-trips/
• https://streamlit.io/
• https://fastapi.tiangolo.com/
• https://www.crewai.com/
• https://docs.ragas.io/en/stable/