SwampSense

Inspiration

Gainesville’s weather and campus microclimates change fast. One walkway floods after a storm, another bakes in the sun, and air quality or noise can vary block by block—but students usually find out only after they’re already there. During Gator Hack V at UF AI Days, we wanted to build something that was unmistakably “swamp” and genuinely useful: a system that treats the campus itself as a sensor network. We were inspired by the idea that almost everyone already carries a phone full of sensors, and that a few cheap IoT devices could fill the gaps. Instead of building another dashboard, we asked: what if the campus could sense itself and warn people in real time?

What it does

SwampSense provides:

A live campus hazard map showing heat, flooding, air quality, noise, and crowd-reported issues.

Real-time alerts when conditions cross safe thresholds, with plain-language summaries generated from structured sensor data.

Safer-route suggestions that avoid flagged zones and prefer shaded, dry, or lower-risk paths.

A facilities dashboard for campus safety and operations teams to see where conditions are changing.

Crowd reports and photos so people can flag flooded paths, broken shade structures, smoke, or other hazards.

Sensor fusion and risk scoring that combines noisy phone data with fixed low-cost sensors instead of trusting any single reading.

The core output is a per-location risk score.

How we built it

We built SwampSense as an end-to-end prototype during the 48-hour hacking period at Newell Hall.

Mobile/web app: A Progressive Web App that uses available phone sensors—GPS, barometer, accelerometer, microphone (opt-in), and camera photos—with manual and crowd-report fallbacks where browser APIs are limited.

Low-cost IoT prototype: An ESP32-based sensor node with a BME280 for temperature/humidity/pressure and a particulate sensor for air quality, used to demonstrate fixed-point sensing.

Backend: A FastAPI service with a PostGIS database for geospatial queries and Redis for short-lived real-time state.

Sensor fusion: We combined fixed and mobile readings with a weighted fusion estimate. AI layer: We used Gemini to turn structured sensor readings and crowd reports into concise, human-readable alerts and summaries—always grounded in templated data so the model could not invent hazards.

Map and alerts: A Mapbox/Leaflet map with risk overlays, alert subscriptions, and a facilities view.

Open source: We kept the project open-source-friendly for the MLH Best Open Source Project category and designed the Gemini integration for MLH Best Use of Gemini.

Challenges we ran into

Sensor noise and calibration: Phone sensors are inconsistent across devices, and low-cost IoT sensors drift. We had to normalize readings and avoid over-trusting any single source.

Sparse data: In a 48-hour hack, we could not deploy hundreds of sensors. We combined a small number of real readings with simulated and crowd-sourced data, and we designed the system to degrade gracefully when data is missing.

Privacy: Microphone and location data are sensitive. We made audio opt-in, processed it locally where possible, and stored aggregated features rather than raw audio.

Real-time geospatial performance: Ingesting, aggregating, and serving location-based risk scores quickly required careful indexing and caching.

AI reliability: We wanted Gemini to help explain alerts without hallucinating hazards. We constrained it with structured inputs and template fallbacks.

Time: Forty-eight hours is short. We prioritized a working end-to-end demo over polishing every edge case.

Accomplishments that we're proud of

Built a working end-to-end prototype: sensor/report ingestion → fusion → risk map → alerts → Gemini summary.

Demonstrated sensor fusion between phone data, a low-cost ESP32 node, and crowd reports.

Created a live hazard map with per-location risk scores and safer-route suggestions.

Added Gemini-generated plain-language alerts grounded in structured data.

Designed for privacy-conscious sensing with opt-in audio and aggregated features.

Shipped an open-source-ready project during Gator Hack V.

What we learned

Sensor fusion is as much about trust and calibration as it is about math.

AI is most useful when it explains structured data rather than replacing it.

Real-time geospatial systems need caching, indexing, and graceful degradation from day one.

Privacy and consent have to be designed in, not bolted on.

A 48-hour hack forces hard prioritization: a narrow, working demo beats a broad, broken platform.

Teamwork and clear interfaces matter as much as code when shipping under time pressure.

What's next for SwampSense

Pilot on UF campus with a small network of fixed sensors and student volunteers.

Improve predictive models for flooding and heat using weather forecasts and historical data.

Add privacy-preserving crowd sensing and better offline support.

Expand safer routing with shade, accessibility, and real-time hazard avoidance.

Partner with campus facilities and safety teams for a real operations dashboard.

Grow the open-source community and document the sensor hardware so other campuses can deploy SwampSense.

Built With

• browsergeolocationapi
• discord
• docker
• fastapi
• figma
• javascript
• notion
• python
• redis