Inspiration

Air quality affects billions daily, yet most tools force users to interpret PM2.5 values, AQI charts, and health categories on their own.
People don’t think in numbers—they think in decisions:

• “Can I go for a run today?”
• “Is it safe for my kids outside?”
• “Should I wear a mask?”

I wanted to build something that makes air quality human, conversational, and instantly understandable.
That’s how AirSight was born.

What It Does

AirSight is an AI-powered air quality assistant that combines:

• Real-time pollution data from 5,000+ global sensors
  
• EPA-standard AQI calculations
  
• Cerebras Llama 3.3 70B for ultra-fast inference
  

You ask a natural-language question like:

“Is it safe to jog in Delhi today?”

AirSight fetches live sensor readings, computes the real AQI, and returns clear, actionable guidance in seconds.

How We Built It

• Built a global city-to-sensor index by scanning all OpenAQ locations and generating a 948-city keyword map
  
• Developed a fast OpenAQ client for real-time PM2.5 retrieval
  
• Implemented EPA PM2.5 → AQI conversion formula for accurate categories
  
• Used Cerebras Llama-3.3-70B for contextual advice grounded in the live data
  
• Designed a clean Gradio UI with debounced autocomplete
  
• Optimized the entire request pipeline to achieve ~1.5s end-to-end latency
  

Challenges We Ran Into

• The OpenAQ city parameter frequently returned incorrect data
  
• Coordinate-based lookups failed with 500 errors for large metros
  
• Filtering 5,000+ sensors per query was too slow for real-time UX
  
• Ensuring the LLM used only real sensor values required careful prompt engineering
  
• Building autocomplete for nearly 1,000 cities without lag required indexing + caching
  

Accomplishments We're Proud Of

• Built a 1ms lookup city index for accurate sensor mapping
  
• Achieved 1.5s total latency combining OpenAQ + LLM inference
  
• Grounded a large LLM reliably in structured environmental data
  
• Created a tool that users immediately found intuitive and useful
  
• Designed a simple architecture that outperformed heavier RAG stacks
  

What We Learned

• Precomputation beats complex runtime pipelines
  
• LLMs become far more reliable when grounded in real numerical context
  
• Environmental APIs require resilient error-handling
  
• Simplicity → faster iteration, easier debugging, and a smoother UX
  

What's Next for AirSight

• Historical AQI trends (“How did Delhi’s air quality change this week?”)
  
• Forecasting with weather + pollution models
  
• Multi-pollutant support (PM10, O₃, NO₂)
  
• Push alerts for threshold exceedances
  
• Mobile apps and home-screen widgets
  
• Community dashboards for runners, schools, and sensitive groups
  

Built With

• asyncio
• cerebras
• fastapi
• gradio
• huggingface
• pydantic
• python