Voyagent

Turn chaos into itineraries with Agentic AI.

Comment

💡 Inspiration

Travel planning is kinda broken. We live in an age of AI, yet we still plan trips by opening 20 browser tabs, checking flights, comparing hotels, reading reviews, looking up weather, and scanning “Top 10” lists. The mental load of stitching all that into one plan is way too much.

So we asked ourselves: what if you had a team of travel experts working for you 24/7? Not just a chatbot that throws generic ideas at you, but a group of specialized agents—a researcher, a local guide, a scheduler, who work together to build a trip that actually makes sense.

That’s how Voyagent started.

🤖 What it does

Voyagent is an agentic AI travel planner. Instead of one big model trying to do everything, we use 8 specialized AI agents that collaborate like a real planning team.

Here’s the pipeline:

  1. Research Agent Finds flights, hotels, and transport options from real sources.

  2. Attractions Agent Pulls top-rated places + hidden gems based on your vibe and interests.

  3. Weather Agent Checks forecast + patterns so outdoor stuff lands on the best days.

  4. Scheduler Agent (the core) The “logic brain.” It fits everything into time blocks, accounts for travel time, opening hours, pacing, and handles “Magic Fill” inserts.

  5. Food Agent Suggests dining that fits the plan (location + time-aware), not random restaurants.

  6. Budget Agent Estimates costs and helps balance splurges vs. savings.

  7. Logistics Agent Makes sure routes and transfers are realistic and you’re not sprinting across a city all day.

  8. Validator Agent Final sanity check: catches overlaps, impossible timing, closed venues, weird travel jumps, etc.

Output: a clean, minute-by-minute itinerary that’s realistic, optimized, and ready to export to your calendar.

🛠️ How we built it

We kept the architecture modular on purpose (agent systems get messy fast if you don’t).

  • Brain: Google Gemini (gemini-2.0-flash) for speed + reasoning
  • Orchestration: custom Python/Django backend as the “Orchestrator” (handles state + agent handoffs)

  • Data layer:

    • Google Places API (real locations, hours, ratings, coords)
    • OpenWeatherMap (forecasts)

  • Frontend: modern JavaScript + CSS with a clean glassy UI

  • Resilience: a stub fallback system so if the AI API is down, the app degrades gracefully instead of breaking

😤 Challenges we ran into

  • The context window trap Agents would forget earlier decisions. We fixed it with a clean context-passing system so each agent gets the exact state it needs.

  • API limits (429 errors) Scaling agents hit rate limits fast. We added caching + exponential backoff retries.

  • Hallucinations vs. reality Models love inventing places. We forced cross-checking with Google Places and added a Validator agent to catch impossible schedules.

  • JSON formatting pain Getting 8 agents to output consistent JSON was… rough. We enforced strict schemas and added a “self-healing” step where the model fixes its own invalid JSON.

🏆 What we’re proud of

  • Magic Fill You can type: “Find a cool jazz bar for tonight” and Voyagent understands context, finds a real place, and drops it into the right open time slot.

  • The UI actually feels premium We moved away from the default “bootstrap look” and built something that feels like a real travel product.

  • A real agent workflow Watching the system work is honestly satisfying—research feeds planning, planning feeds scheduling, scheduling gets validated.

🧠 What we learned

  • Specialization beats one giant prompt Smaller focused agents > one mega prompt.

  • Latency matters People hate waiting, so we stream progress like: “🔍 Researching…” → “✅ Validated” to keep it feeling fast.

  • Data validity is everything A perfect-looking plan is useless if the museum is closed. Real APIs + validation are non-negotiable.

🚀 Future scope (what’s next)

Short term

  • Booking integration Turn “Book now” into real hotel/flight/activity bookings.

  • Group trips Let friends vote on activities and build a shared plan without chaos.

  • Mobile app A React Native version for on-the-go edits.

Longer term

  • Hyper-personalization Opt-in connections (Spotify/Instagram/etc.) to suggest places based on actual taste, not generic profiles.

  • Real-time replanning If weather changes or something closes, Voyagent reshuffles your day automatically.

  • Smarter budget optimization Not just estimates—real tradeoffs like: “keep it under $2k” and the itinerary adjusts.

  • Trip companion mode From planning → booking → day-of navigation → post-trip recap (places visited, spend summary, shareable itinerary).

Built With

Share this project:

Updates