Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Atmos

Hyperlocal air quality forecasting platform for India.

Atmos is an end-to-end machine learning and AI system built to close a critical gap in India's air quality monitoring — national AQI stations are too sparse to give people block-level visibility into pollution, which is the resolution that actually matters for personal health decisions.

Built as a college major project, targeting production-grade AI/ML engineering practices from data ingestion through to LLM-powered user-facing insights.


Overview

Atmos ingests live data from 250+ AQI monitoring stations across India, along with satellite fire detection feeds and weather data, into a structured data pipeline. It trains spatiotemporal machine learning models to forecast AQI up to 72 hours ahead at a hyperlocal resolution, and uses the Claude API to translate raw forecasts into plain-English explanations and personalized health advisories.

The project is being built in three phases:

  • Phase 1 — Data Pipeline (complete)
  • 🔄 Phase 2 — ML Modeling (in progress)
  • Phase 3 — Serving + LLM Layer (planned)

Architecture

Live Data Sources (AQI stations, satellite, weather)
              │
              ▼
     Apache Airflow (orchestration)
              │
              ▼
   TimescaleDB (bronze → silver → gold)
              │
              ▼
  Spatiotemporal ML Models (LightGBM, XGBoost)
              │
              ▼
       FastAPI Backend
              │
              ▼
   Claude API (explanations + health advisories)
              │
              ▼
          Frontend

Data Pipeline

A multi-source ingestion pipeline built with Apache Airflow pulls in:

  • Live readings from 250+ AQI monitoring stations
  • Satellite-based fire detection data
  • Meteorological / weather data

Data flows through a bronze / silver / gold architecture on TimescaleDB, designed to handle partial data gaps and inconsistent update frequencies across sources without breaking downstream model training.

ML Modeling

Spatiotemporal models built with LightGBM and XGBoost, using engineered features that capture both time and spatial pollution dispersion patterns — standard time-series models don't account for how pollution physically moves across a region, so this required custom feature engineering beyond off-the-shelf approaches.

Target: AQI forecasts up to 72 hours ahead at a resolution finer than what national monitoring stations provide.

LLM Layer

The Claude API is integrated to:

  • Generate plain-English explanations for why a pollution spike is happening
  • Deliver personalized health advisories based on individual user health profiles

This turns a raw forecast number into something a non-technical user can actually act on.


Tech Stack

Layer Technology
Orchestration Apache Airflow
Database TimescaleDB
ML Modeling LightGBM, XGBoost, Python
Backend FastAPI
LLM Claude API
Data Sources AQI monitoring APIs, satellite fire detection, weather APIs

Status

This is an actively developed project. The data pipeline (Phase 1) is complete and stable. ML modeling (Phase 2) is in progress. Contributions, feedback, and issues are welcome.


Author

Gautam Sharma B.E. Information Science and Engineering, Bangalore Institute of Technology

About

Hyperlocal air quality forecasting with LLM-powered source attribution and health advisories.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages