LLM-Powered IoT Device Monitoring Platform

Problem

IoT device fleets generate massive volumes of telemetry data — GPS signals, camera feeds, CPU metrics, temperature readings, and error logs — all stored across Elasticsearch indices. Operations teams were spending significant time manually querying dashboards and sifting through logs to identify issues, often missing subtle cross-signal patterns that indicated impending device failures.

Solution

I built an LLM-powered monitoring platform deployed on Google Cloud Platform that automates telemetry analysis through intelligent tool orchestration:

• LLM Function Calling Orchestration: Designed the core agent loop where the LLM dynamically selects and invokes the right tools based on the analysis context — deciding which data to fetch, how to process it, and what to report
• Tool Suite: Built a set of callable tools including Elasticsearch data retrieval (querying device metrics across time ranges), data cleaning and aggregation pipelines, anomaly detection routines, and report formatting functions
• Context Management: Implemented structured context passing so the LLM maintains awareness of fleet-wide state across multi-step analysis workflows — correlating anomalies across different devices and signal types
• Automated Daily Reports: The system generates comprehensive reports with natural language explanations of device behavior patterns, flagged anomalies with probable root causes, and prioritized recommendations

Architecture

Elasticsearch (telemetry data)
        |
  Tool Functions ── Data retrieval
        |          ── Data cleaning
        |          ── Anomaly detection
        |          ── Report formatting
        |
  LLM Orchestrator (function calling)
        |── Selects tools & parameters
        |── Manages multi-step context
        |── Generates natural language insights
        |
  Daily Analysis Report

Impact

• Replaced manual daily telemetry review with automated, LLM-orchestrated analysis
• Enabled operations teams to focus on acting on insights rather than searching for them
• Improved early detection of device issues through multi-signal pattern correlation that manual review consistently missed

Technical Highlights

• LLM function calling for dynamic tool selection — the model decides which Elasticsearch queries to run and how to interpret results
• Robust parameter extraction and validation to ensure correct tool invocation
• Multi-step reasoning with context accumulation across sequential tool calls
• Deployed as a managed service on GCP with scheduled execution
• Graceful error handling when tools return unexpected data or LLM hallucination detection in generated reports