Multi-agent LLM web app that helps newcomers find the right Sydney suburb using AI, combining Reddit sentiment, crime statistics, transport data, and civic amenities into a conversational assistant. Won the Best AI LLM Project Award at UTS 2026.
Built for the Applied NLP subject at UTS as a team of five, Sydney Liveability AI is a production-grade AI system that answers natural-language questions about where to live in Sydney. The user defines a personalised weight profile across five dimensions (safety, transport, lifestyle, affordability, proximity to CBD) through a conversational onboarding flow.
The system then ranks all 656 Greater Sydney suburbs, displays the top 5 on an interactive map, and lets the user explore each suburb through a chat grounded in cited evidence. The backend is a multi-agent CrewAI pipeline built on FastAPI: a Router agent classifies each query and dispatches it to specialist agents (Crime, Sentiment, GIS, Comparator), and a Synthesiser agent assembles the final response with auditable evidence traces.
The sentiment pipeline combines BART-MNLI zero-shot classification, DeBERTa-v3 ABSA, and GoEmotions emotion profiling over 20,423 Reddit posts across 563 suburbs. Structured data (BOCSAR crime stats, TfNSW GTFS transport feeds, OSM amenities, ABS housing data) lives in PostgreSQL via Supabase. Vector data (Reddit posts, PDF narratives) lives in ChromaDB with MiniLM-L6-v2 embeddings.
The frontend is built in Next.js with an onboarding panel, interactive map, chat interface, and evidence drawer. Architecture designed and led by Nelkit Chavez, including FastAPI endpoints, Synthesiser agent, scoring formula, ArcGIS ingestion, and full frontend. Won the Best AI LLM Project Award in the UTS subject-wide competition.
