SpartaSafe

SpartaSafe
Landing page
Live map feed

Inspiration

Campus crime data is often technically accurate but difficult for students to interpret in real time. Incident logs, police feeds, and alerts are usually verbose, inconsistent, or written for compliance rather than everyday decision-making. We were inspired to build SpartaSafe after noticing how hard it is for students to quickly answer simple questions like “Is it safe to walk here right now?” or “Has anything happened near my building recently?” without digging through dense reports or multiple sources.

Our goal was to make campus safety information clear, calm, and aactionable, without sensationalizing crime or overwhelming users.

What it does

SpartaSafe is an AI-powered campus safety assistant that summarizes recent crime and safety incidents around campus in a human-readable way. Users can:

View recent incidents with clear summaries and trends
Check safety scores for specific campus locations
Receive calm safety guidance when no incidents are reported
Ask follow-up questions to filter by time window, location, or source

Instead of dumping raw records, SpartaSafe focuses on context, patterns, consistent performance, and practical next steps.

How we built it

We built SpartaSafe as a full-stack web application with:

PostgreSQL, Prisma for structured incident storage
Supabase for hosting the database
LangChain to orchestrate an AI agent with tools
Local map using Leaflet.js
Gemini models to generate concise, structured summaries
A custom safety prompt enforcing consistent Markdown formatting
A React/Node.js frontend with a chat-style interface
WebSockets to reload the page in real-time without manual inteference

Incident data is normalized server-side into display-ready fields (local time, human-friendly locations, translated dispositions) before being passed to the AI. This prevents “debug dump” outputs and ensures the model focuses on interpretation rather than formatting.

Challenges we ran into

One of the biggest challenges was balancing accuracy with readability. Raw crime data includes internal IDs, coordinates, timestamps in UTC, and coded dispositions. Exposing this directly made outputs confusing and overwhelming.

We also ran into:

Performance issues caused by rebuilding the AI agent on every request
Markdown formatting being flattened by frontend CSS resets
Bullet points and spacing disappearing due to global style overrides
Deciding when the AI should access the database vs. respond with general safety guidance

Solving these required careful coordination between backend tooling, prompt design, and frontend rendering.

Accomplishments that we're proud of

Designing a strict prompt and tool contract that consistently produces clean, readable outputs
Transforming raw incident data into human-centered safety summaries
Eliminating noisy internal fields without losing important context
Achieving fast response times by caching the AI agent and limiting tool calls
Creating a calm UX that informs without causing unnecessary alarm

What we learned

We learned that AI systems are most effective when:

Tools return exactly the data users should see, not everything available
Formatting and presentation are treated as first-class design problems
Frontend, backend, and prompt design are aligned as one system

We also gained hands-on experience with modern AI agent patterns, database introspection, schema migrations, and production-grade UI debugging.

What's next for SpartaSafe

Next, we want to expand SpartaSafe by:

Supporting multiple campuses with configurable data sources
Introducing confidence indicators to distinguish official reports from community submissions
Allowing users to subscribe to personalized safety alerts
Improving trend analysis using lightweight statistical summaries (e.g., frequency changes over time)

Ultimately, we hope SpartaSafe can become a trusted, everyday tool for helping students stay informed and feel safer on campus.