Inspiration
The travel industry faces a massive imbalance: a handful of popular destinations suffer from severe overtourism and ecological degradation, while countless breathtaking, offbeat locations remain undiscovered. Local communities in these hidden areas miss out on the economic benefits of tourism. We realized that traditional travel apps just push the same top 10 tourist traps. We built Project Sentinel to change this by using AI to autonomously uncover hidden gems and dynamically route travelers to sustainable, offbeat destinations.
What it does
Project Sentinel is an autonomous, AI-driven travel concierge that generates highly personalized, day-by-day itineraries focused on sustainable tourism. Instead of relying on static, generic databases, Sentinel features a live, autonomous data pipeline. It actively mines YouTube travel vlogs to find verified offbeat locations, parses the data, and stores it in our live database. When a user asks the Next.js frontend for a trip (e.g., "A quiet nature escape to avoid the crowds"), our Gemini-powered RAG (Retrieval-Augmented Generation) engine cross-references the user's vibe with our hidden gems database. It then outputs a conversational, empathetic response alongside a beautiful, interactive visual slide deck featuring dynamic maps and imagery.
How we built it
Our architecture is split into a seamless, two-part ecosystem:
- The Autonomous Scraper (Data Pipeline): We built a Python script that runs continuously via GitHub Actions. It uses the
youtube-transcript-apito scrape travel vlog transcripts, feeds that raw text into Google Gemini 3.5 Flash-Lite to extract geospatial data and location vibes, and pushes the cleaned JSON to a Google Sheets Webhook. - The RAG Frontend (User Interface): We built the client interface using Next.js (App Router), React, and Tailwind CSS. When a user submits a prompt, the backend fetches the live data from Google Sheets, injects it into a strict Gemini prompt template, and generates a structured JSON itinerary.
- Dynamic Visuals: We integrated the Unsplash API via a custom route to dynamically fetch unique landscape photography based on the AI's generated locations, mapping them directly to interactive itinerary cards.
⚠️ Challenges we ran into
- Hallucination Control in RAG: Ensuring Gemini strictly prioritized our scraped database of hidden gems over its pre-trained knowledge of generic tourist spots required heavy prompt engineering and JSON schema enforcement.
- Bridge Architecture: Figuring out how to connect a completely separate Python scraping environment to a Next.js frontend. We solved this creatively by using Google Apps Script and Google Sheets as a lightweight, live Webhook database.
- Dynamic Image Matching: Offbeat locations often lack direct image matches on stock APIs. We had to build a custom fallback and randomization endpoint so the UI always remained beautiful and engaging, even when specific coordinates were too obscure.
Accomplishments that we're proud of
- Successfully engineering a fully autonomous pipeline that goes from raw YouTube video transcripts all the way to a polished, interactive UI without human intervention.
- Achieving incredibly fast and accurate JSON structuring using Gemini 3.5 Flash-Lite.
- Designing a sleek, professional UI (featuring a cohesive White, Olive, and Gold aesthetic) that feels like a premium concierge service rather than a standard chatbot.
What we learned
- Advanced prompt engineering and JSON schema enforcement with the Gemini API.
- How to build lightweight, zero-cost data pipelines using Google Sheets Webhooks.
- Handling real-time UI state management for interactive, multi-modal chat responses in Next.js.
What's next for Project Sentinel
While our initial pilot dataset focused on the Jammu & Kashmir region as a proof-of-concept for navigating mountainous terrains, Project Sentinel is built to scale globally. Next, we plan to:
- Scale our Python scraper to analyze travel vlogs worldwide.
- Integrate real-time weather APIs to dynamically alter itineraries (e.g., suggesting indoor hidden gems if rain is detected).
- Add direct local-business booking integrations to ensure tourist spending goes straight to the local economy.
Built With
- gemini
- github-actions
- google-sheets
- next.js
- python
- react
- tailwind-css
- typescript