air_quality

🌫️ Delhi AQI Intelligence — Specialized Air Quality Analyst Model Problem Statement Delhi is one of the most polluted cities in the world, yet most residents lack accessible tools to understand what the pollution data actually means for their health. Raw AQI datasets contain technical pollutant readings (PM2.5, PM10, CO, NO₂, O₃, SO₂, NH3) that are meaningless to the average person without domain expertise. General-purpose LLMs can answer broad questions, but they lack the ability to:

Reason over a specific, real-time dataset with hourly granularity Apply India-specific CPCB AQI standards consistently across responses Provide grounded health advisories tied to actual recorded values Detect temporal pollution spikes and explain their likely causes

Solution A specialized fine-tuned model trained on Delhi's hourly AQI dataset (January 2023), grounded in CPCB and WHO pollutant thresholds. The model acts as a domain expert that can interpret any pollutant reading, flag health risks, and explain trends in plain language. Why a Specialized Model vs. a General LLM? FactorGeneral LLMThis Specialized ModelData groundingHallucinates valuesAnswers from actual CSV dataAQI standardsGeneric/globalIndia CPCB-specific thresholdsHealth advisoriesVagueTied to recorded pollutant levelsPollutant expertiseSurface-levelDeep: CO, NO, NO₂, O₃, SO₂, PM2.5, PM10, NH3Temporal reasoningNoneHourly trend detection A general LLM asked "Was Delhi's air safe on Jan 2nd at 3AM?" would guess. This model knows PM2.5 was 940.49 µg/m³ at that hour — 37× the safe limit — and responds accordingly. Dataset

Source: Delhi Hourly AQI Readings, January 2023 Pollutants tracked: CO, NO, NO₂, O₃, SO₂, PM2.5, PM10, NH3 Granularity: Hourly readings across 20+ days Key finding: PM2.5 regularly exceeded 800–1,100 µg/m³ during peak hours (WHO safe limit: 25 µg/m³)

Model Capabilities

Interpret any pollutant value against CPCB AQI categories (Good → Severe) Identify the most dangerous hours and days in the dataset Explain health risks for sensitive groups (children, elderly, respiratory patients) Compare multiple pollutants simultaneously Detect overnight pollution buildup patterns common in Delhi winters

Base Architecture Gemma 3 (270M) — fine-tuned for low-latency, domain-specific inference

Built With

• csv
• python