PyroGuard

PyroGuard Logo

Inspiration PyroGuard Sentinel was inspired by both personal loss and professional experience. Having lived most of my life in West Maui, I saw firsthand how devastating the Lahaina wildfire was for my community. In the weeks that followed, I worked as a paid disaster response contractor for the American Red Cross, first on the ground with survivors and later reconciling critical data for the organization. That experience exposed how reactive and fragmented wildfire response systems often are. It became clear that more intelligent, transparent, and rapid tools are needed to help communities anticipate and respond to fire risk before it’s too late.

What it does PyroGuard Sentinel is an AI-powered wildfire early warning system that performs fast, transparent risk analysis based on real-world satellite, weather, and infrastructure data. It identifies dry vegetation using image analysis, factors in weather patterns like wind and humidity, and checks nearby power-line density using geospatial data. All of this is orchestrated through a Model Context Protocol (MCP) pipeline that explains each step in plain language. Once complete, the system auto-generates structured Jira incident reports, giving emergency managers or first responders something they can act on immediately.

How I built it The core image data comes from the open Sentinel-2 imagery stored in AWS’s sentinel-cogs S3 bucket. Clarifai’s NDVI model was originally selected to evaluate vegetation dryness, but due to limitations in accessing the right pretrained model during the hackathon, I added fallback logic using Anthropic's Vision API. For power-line detection, I used Overpass API with a 500-meter radius. NOAA provided the latest wind, humidity, and temperature data via weather.gov. All these data sources flow through a FastAPI backend, orchestrated using Inngest for retry logic and Operant AI for step-by-step timeline visibility. The frontend is built in Next.js and deployed via Vercel, while the backend and worker processes are hosted on Render. Tickets are filed automatically through Make.com into a connected Jira board. The entire system is end-to-end automated and designed to deliver results in under 20 seconds.

Challenges I ran into One major hurdle was Clarifai access I wasn’t able to get a suitable vision model added in time, which is why I included Anthropic’s API as a fallback to ensure results stayed consistent. I also ran into challenges around data resolution and model accuracy; some vision outputs were ambiguous or lacked confidence. Addressing this in a production setting would require access to more sophisticated computer vision tools or better raw data inputs. Coordinating latency, retries, and streamable UI updates within a tight 20-second budget added pressure but ultimately shaped a stronger architecture.

Accomplishments that I'm proud of I’m proud of building something that reflects both personal motivation and real technical value. PyroGuard brings together five sponsor tools and a wide array of data sources into one fast, seamless demo that’s not only functional but deeply meaningful to me. Every part of the system imagery, reasoning, analysis, and reporting is visible and explainable. The result is a fully automated agent that actually helps solve a real-world problem I’ve lived through.

What I learned This project taught me how to work with open satellite data at scale and how to pair it with AI reasoning to create contextual analysis. I learned how to orchestrate a Model Context Protocol flow with multiple fallback conditions, how to build retry-safe pipelines with Inngest, and how to streamline complex information into a user interface that stays responsive. More than anything, I learned that combining personal experience with technical skill can lead to something powerful and relevant.

What's next for PyroGuard I plan to evolve PyroGuard into a robust alerting platform for municipal and state-level emergency teams. I want to improve its vegetation detection accuracy, bring in better fire behavior prediction models, and explore integrations with FEMA or utility providers. Long term, the goal is to make PyroGuard a plug-and-play layer that can slot into any emergency response workflow, delivering explainable, real-time fire risk assessments anywhere in the world.