Beacon Shield

Inspiration

I grew up in regions where wildfires and flash floods weren’t hypothetical—they were seasonal. During crises, I often watched neighbors scramble for trustworthy information. LuminHacks’ “anyone can build with AI” challenge inspired me to prove that a single maker, using local-first AI, could build a mission-critical command center for communities without relying on big cloud stacks.

What it does

BeaconShield is my local-first crisis orchestration platform. It simulates wildfire, flood, and extreme-heat scenarios, analyzes risk, and uses an on-device t5-small model to generate tailored action plans, resident communications, and volunteer assignments. Everything runs in a single-page dashboard so a community leader can coordinate directly from their laptop—even offline after initial setup.

How I built it

I paired a FastAPI backend with deterministic simulators to mimic hazard feeds, added lightweight analytics to turn raw metrics into risk signals, and wired a structured prompt into t5-small to emit JSON action plans. The frontend is a vanilla HTML/JS app that streams API responses into risk badges, checklists, and volunteer rosters. Once dependencies download, the stack runs entirely on my machine.

Challenges I ran into

• Local-only AI: Keeping inference offline meant juggling model size, performance, and caching the weights ahead of time.
• Structured outputs: Coaxing t5-small to consistently emit parseable JSON required meticulous prompt tuning and fallback logic.
• Believable simulations: With no live feeds, I had to design stochastic yet deterministic hazard simulators that still felt realistic for demos.
• Solo scope control: As a team of one, I had to prioritize ruthlessly to ship multi-hazard coverage without sacrificing polish.

Accomplishments that I'm proud of

• Delivering a complete, test-covered incident platform that never needs to hit the cloud after install.
• Getting a compact model to produce actionable, role-specific checklists and community messages from synthetic data.
• Presenting a cohesive dashboard that communicates complex crisis insights in an accessible, judge-ready format.

What I learned

• Careful prompt engineering plus deterministic analytics can make small models behave like structured planners.
• Transparency—showing raw metrics next to AI-generated actions—builds user trust in high-stakes situations.
• Offline-first architecture is absolutely achievable solo when you pre-stage large dependencies and keep your toolchain lean.

What's next for BeaconShield

• Integrate live weather APIs, LoRa sensor gateways, or drone imagery to move beyond simulations.
• Add collaboration features such as shared timelines, acknowledgement flows, and after-action logging.
• Fine-tune or distill domain-specific models for wildfire or flood playbooks while keeping inference local.
• Co-design with emergency managers to validate workflows, accessibility, and compliance needs before piloting in real communities.

Built With

• css3
• fastapi
• html5
• huggingface
• numpy
• pandas
• pydantic
• python
• pytorch
• uvicorn