Smart Wardrobe Suggestion

Architecture – LLM Weather-Based Clothing Suggestions

This document describes the architecture of the LLM-powered weather-based clothing suggestion system built using Apache Airflow.

The system fetches short-term weather forecasts, reasons over them using a language model agent, and delivers concise clothing recommendations to the user via push notifications.

High-Level Overview

The solution consists of four major layers:

1. Orchestration Layer (Apache Airflow)

2. Data Acquisition Layer (Weather API)

3. Decision Intelligence Layer (LLM Agent)

4. Notification Layer (PushBullet)

Each layer is isolated using custom Airflow operators, making the system modular and extensible.

Workflow Summary

• The DAG runs twice daily (07:00 & 14:00, Europe/London).

• Weather data is retrieved for Bromley, UK.

• A constrained LLM agent interprets the forecast and produces clothing advice.

• The final recommendation is delivered directly to the user’s device.

Detailed Workflow

1. Workflow Orchestration

• DAG ID: dag_llm_weather_suggestions

• Schedule: 0 7,14 * * *

• Timezone: Europe/London

• Catchup: Disabled

The Airflow Scheduler triggers the DAG based on the defined cron schedule.

2. Weather Retrieval (get_weather)

Operator: WeatherApiOperator

Responsibility:

• Queries an external Weather API.

• Retrieves forecast data for the next hour.

Inputs:

• Latitude: 51.406

• Longitude: 0.015

• Forecast window: 1 hour

Outputs:

• Structured weather forecast JSON (pushed to XCom).

This task performs no interpretation, only data collection.

3. Clothing Recommendation Engine (get_clothing_suggestions)

Operator: LlmClothingAgentOperator
Base Class: BaseOperator
Template Fields: weather_conditions