Pyre Project Experience

Experience Building Pyre

Jun 10, 2025 • 13 min read

Pyre: AI-Powered Wildfire Prediction System - A Journey Through Real-World Problem Solving

Executive Summary

Pyre is an AI-powered wildfire monitoring platform designed to enhance emergency preparedness and real-time response. It integrates environmental data, user-reported hazards, and live fire incidents into a seamless single-page application tailored for emergency scenarios. Throughout this project, I led engineering efforts, managed team dynamics, and pushed for iterative improvements based on both user feedback and field testing.

Project Vision & Impact

Wildfires are fast, brutal, and often under-monitored. Pyre closes the gap between scattered environmental data and actionable insights. It empowers residents, firefighters, and local governments to make critical decisions during wildfire threats with a clear, unified platform.

Core Challenges & Solutions

1. Unclear Project Value Proposition

Problem: Stakeholders weren’t sold on our pitch — too much tech jargon, not enough real-world relevance.

Why: We buried the lede. Instead of showing how Pyre could save lives, we showcased technical features without clear user benefit.

Fix:

  • Led with wildfire casualty stats and the problem statement.
  • Developed user personas for first responders, local officials, and evacuees.
  • Walked through realistic emergency workflows.
  • Added demo quotes from San Diego fire department contacts who beta-tested the prototype.

Result: Turned confusion into support — got buy-in from emergency planners and local organizations.

2. Scattered User Experience

Problem: Tools were fragmented across multiple pages — a nightmare when seconds matter.

Finding: Testing showed it took users ~45 seconds to access evacuation info. That’s a death sentence in a fast-moving fire.

Fix:

  • Merged everything into a unified single-page app.
  • Added modals for fast switching between tools.
  • Introduced a high-contrast emergency mode and keyboard navigation.
  • Slashed task completion time to 8 seconds via design overhauls and user testing.

Impact: 75% more engagement, vastly reduced abandonment under stress.

3. Missing AI Assistant Functionality

Problem: Users struggled with tool navigation and didn’t know where to start during emergencies.

What We Built Instead of ML:

  • I integrated Google’s Gemini AI using the Gemini API key I registered and secured through service configuration.
  • Created an in-app chatbot using modal architecture so it could live on every screen.
  • Trained the assistant with Pyre’s feature set and basic fire safety protocols.

Functionality:

  • Users can ask where to evacuate, how to report a fire, or get a route in real-time.
  • Gemini handles natural language requests and guides them to the right tool instantly.

Why It Worked:

  • Replaced traditional machine learning models with a much more adaptable assistant.
  • 90%+ of testers preferred the AI assistant for navigation over traditional menus.

4. Team Coordination Woes

Problem: A six-person team, two subsystems (fire + earthquake), one tangled mess of integrations.

Symptoms: Merge conflicts, duplicated features, API miscommunication — all the classics.

My Fix:

  • Co-led agile planning with teammate Pranav.
  • Implemented strict API documentation protocols.
  • Added weekly syncs between fire and quake teams.
  • Built integration testing pipelines for all REST endpoints.

Result: Integration bugs dropped 80%. Dev speed jumped 40%.

5. Performance & Scalability

Problem: Our first test deployment loaded like it was stuck in molasses. 30+ seconds to see anything useful.

Analysis: Large datasets, excessive API calls, zero caching. Basically, rookie mistakes.

My Fixes:

  • Implemented lazy loading + pagination.
  • Cached Gemini API responses and weather data to avoid redundant requests.
  • Preprocessed incoming incident feeds before frontend rendering.
  • Set up CDN distribution for static content.

Result: Load time dropped 95%. We hit 99.9% uptime during simulated emergency testing.

6. Community Reporting Wasn’t Trusted

Problem: How do you crowdsource fire hazard reports without letting trolls or panicked misreports hijack the system?

Solution:

  • Designed a robust reporting tool with field validation, geotagging, and image uploads.
  • Built an admin dashboard for verifying incoming reports.
  • Reports flow through: Pending → Verified → Resolved.

Tech Stack: Flask-RESTful APIs with role-based permissions and rate limiting. Geo-indexed PostgreSQL database for spatial queries.

Impact: Emergency team feedback showed 40% faster hazard response compared to traditional call-in reports.

7. Emergency Communication Delays

Problem: If people don’t get the message in time, everything else we built is meaningless.

Solution:

  • Integrated Twilio API for automated voice alerts.
  • Developed location-based alert logic — only notify those at real risk.
  • Scaled to simulate high-load emergencies, with 95% of calls completed within 3 minutes.

Future-Ready: Framework supports SMS, push notifications, and email — ready to scale with user needs.

Leadership & Deployment

Engineering Lead

  • Spearheaded fire-side subsystem development.
  • Oversaw Gemini AI integration, frontend redesign, and system unification.

Agile Workflow Setup

  • Ran 2-week sprints with velocity tracking.
  • Maintained GitHub Project boards and ran retrospectives.
  • Authored detailed user stories from multiple stakeholder perspectives.

Deployment & DevOps

  • Productionized Flask backend using Gunicorn + Nginx.
  • Optimized PostgreSQL with indexing and connection pooling.
  • Configured logging and monitoring using Grafana + Prometheus.
  • Seamlessly deployed under Open Coding Society’s platform umbrella.

Skills Highlighted

  • Frontend: React.js, ES6, modal architecture, accessibility
  • Backend: Flask, REST API design, authentication, rate limiting
  • DevOps: Gunicorn, Nginx, CDN, caching, uptime monitoring
  • Integration: Gemini AI, Twilio, Google Maps, NASA FIRMS, San Diego Police feeds
  • Teamwork: Agile leadership, stakeholder feedback, sprint management
  • UX/UI: Single-page architecture, chatbot design, task optimization

Final Reflection

Pyre wasn’t just a project — it was a stress test of what I could build, lead, and fix under pressure. While others went the traditional machine learning route, I doubled down on user-centric design and practical AI integration. We made an intuitive, production-ready tool that’s already drawing real-world interest.

In a crisis, people don’t need more dashboards. They need clarity. Pyre delivers that — fast, clean, and backed by engineering that can hold under fire.