🌌 01 · Main Home & Exploration Studio
An autonomous multi-agent travel system that builds spatially coherent day-by-day itineraries. It combines LangGraph state machines with K-means++ coordinate clustering, Reddit community sentiment scoring, and an interactive drag-and-drop timeline with Playwright PDF export.
Standard AI travel planners are simple wrappers around single LLM prompts. They often hallucinate attractions, generate geographically chaotic routes (causing zig-zag transit across cities), push generic tourist traps, and lack persistent state for iterative editing.
WanderAI is a multi-agent state machine built with LangGraph. Four specialized agents (Intake, Ranker, Planner, Editor) parse requirements with dynamic clarifying questions, surface authentic hidden gems using sentiment-scored Reddit discussions, group attractions using K-means++ geographic clustering, and provide interactive timeline drag-and-drop editing with instant Playwright PDF export.
Detects brief prompts (e.g., "3 days in Kyoto") and leverages Gemini 2.0/3.5 Flash to generate context-specific preference chips (vibe, dietary needs, pace) before execution.
Cross-references Google Places popularity with Reddit and travel blog sentiment via VADER intensity and Tavily search to surface verified, under-the-radar local gems.
Partitions attractions into geographically coherent daily clusters with maximally-spread centroid initialization and regional multi-zone dispersion for wide destinations (Goa, Rajasthan, Bali).
Autonomous Editor Agent classifies user feedback (swap_stop, remove_stop, adjust_pace, tell_me_more) and performs live state patching with transit recalculations.
Interactive dual view (Cards 🗂️ vs Timeline ⏱️) with concrete time-slot scheduling (09:00 AM – 10:15 AM), drag handles ⋮⋮, and live sequential transit recalculations.
Zero-auth image pipeline: Wikipedia REST Summary → Wikipedia OpenSearch → Generator PageImages → Wikimedia Commons direct search, ensuring 100% verified real landmark photos.
Custom atmospheric travel narration model fine-tuned on Llama 3.2 3B Instruct via Unsloth/PEFT with GGUF export and zero-latency local Ollama inference fallback.
Headless Playwright PDF brochure generation with Google Fonts & hero covers, RFC 5545 .ics calendar sync, and public shareable /trip/[slug] routes.
Activity- and climate-aware packing checklist generator with interactive checkboxes, live Open-Meteo forecasts, and Google Maps turn-by-turn navigation deep links.
Follow the multi-agent travel planning journey from initial landing and preference intake to real-time agent reasoning and final interactive map generation.
Nocturnal Voyager constellation landing interface with real-time 6-agent status badge and quick query launcher.
Instant curated inspiration cards for Goa, Rajasthan, Lisbon, and Kyoto with pre-tuned regional parameters.
Natural language trip prompt input with one-click dietary bias filters (Vegan, Halal, Vegetarian, Gluten-Free, Jain).
Option-based builder configuring duration, travel vibes, sightseeing pacing, budget tiers, and traveler group types.
Live LangGraph execution stream displaying intake_agent extraction, ranker scouting, and planner_agent structuring.
Day-by-day tabs with geocoded attraction pins, live route polylines, weather summaries, and export capabilities.
Detailed timeline stop cards featuring local historical context, estimated costs, walk times, and swap/remove controls.
hidden_gem_score = clamp(
(min(mention_count, 20) / 20) * ((avg_sentiment + 1) / 2) # Niche community signal
+ (0.2 if source_diversity >= 2 else 0) # Cross-platform bonus (Reddit + Tavily)
- log(review_count + 1) / log(max_review_count + 1), # Log-normalized popularity penalty
min=0.0, max=1.0
)intake_agent.py — Extracts destination, dates, budget; invokes dynamic clarifying questionsranker_agent.py — Concurrently pulls OpenTripMap POIs and Reddit/Tavily community gems, blending by niche weightplanner_agent.py — Performs K-means++ clustering (k = num_days), assigns day themes, and attaches Open-Meteo weathereditor_agent.py — Handles user adjustments (swap, remove, adjust pace) and updates itinerary state with SQLite checkpointsollama_narrator.py — Connects to local LoRA fine-tuned Llama 3.2 3B model for atmospheric, non-cliché stop descriptionsNext.js 16 (React 19) — Nocturnal Voyager glassmorphic design system with Aurora animated background meshMapView.tsx — Leaflet interactive map with CartoDB Dark Matter tiles, marker popups, and route polylinespdf_generator.py — Headless Playwright service rendering print-ready multi-page travel brochureshistory_store.py — SQLite trip database with dual-layer auto-saving and slide-over history drawercurrency.ts — Smart currency engine rendering ₹ INR for Indian destinations and $ USD internationallyLangGraph StateGraph, FastAPI (async), SQLite checkpoints, Pydantic v2 schemas, Python 3.12.
Next.js 16 App Router, TypeScript, React 19, Leaflet, CartoDB Dark Matter, HTML5 Canvas live wallpaper.
Gemini 3.5 Flash (themes & narrations), Groq Qwen 3.6 27B (fast slot-filling), Fine-Tuned Llama 3.2 3B (LoRA / Ollama).
ChromaDB embedded vector store (v7 cache), OpenTripMap, Nominatim, Reddit JSON API, Tavily, VADER.