About SafeHer

What this build is about

SafeHer is an AI-powered safety platform built for Girls Hack Day, Delhi (Problem Statement 12: AI-powered system that identifies unsafe locations based on community reports and public data).

It answers a simple question:

"Safety isn't just about where crime happened yesterday — it's about understanding what's happening around you today."

Instead of publishing static crime statistics, SafeHer fuses three signals into one live, explainable safety heatmap:

• Historical crime data — NCRB / data.gov.in district and police-station-level data
• Live news intelligence — an AI-filtered GDELT pipeline that classifies and dedupes real-world news in real time
• Community reports — hyperlocal, source-cited incident reports submitted by users (text, photo, GPS pin, category)

Every report is analyzed by Google Gemini (summarized, categorized, severity-scored 1–5, spam-checked via embedding similarity) and automatically corroborated — 3+ independent reports in the same geohash cell within 30 days upgrades it from unverified to community-corroborated, with zero manual moderation.

Data honesty is a design principle. Public crime data is district/station-level, never pin-precise. SafeHer never pretends otherwise — precision comes from fusing that coarse baseline with live, cited, hyperlocal reports.

Why I built it

• The problem is personal and real. Public safety apps stop at "view a map of old stats." By the time government data is published, it's often months old. SafeHer exists to close the gap between what happened and what's happening now.
• Current tools don't explain themselves. A "red zone" with no justification helps nobody. SafeHer is built so every hotspot shows exactly which recent incidents and news events contributed to it.
• Current tools are passive. SafeHer is designed to eventually act — with Smart SOS, , and emergency audio recording (all roadmap items), it moves from "telling you a place is unsafe" to "keeping you safe while you're in it."
• A fun engineering challenge. Combining structured crime data, unstructured live news, community crowdsourcing, embeddings, geohash corroboration, and a risk-scoring engine into one coherent system — in hackathon time — was genuinely exciting to build.

How it can be useful for others

• For women and commuters — street-level, up-to-date safety awareness before you walk, travel, or pick a commute route; a one-tap record of incidents that genuinely happened near you.
• For city planners and NGOs — a crowdsourced, timestamped signal of safety hot-spots that complements (and corrects for the lag of) official crime statistics.
• For journalists and researchers — a publicly visible trail of source-cited community reports and AI-classified news events that can be audited.
• For builders — the repo is a complete, modern monorepo template (Turborepo, Next.js, Hono, Bun, Prisma, PostgreSQL with PostGIS + pgvector, Better Auth, shadcn/ui) that demonstrates how to wire Gemini structured JSON output, embeddings, and semantic similarity into a production-grade full-stack app.
• For the next hackathon team — a realistic, honest take on "AI for public safety" that tackles the hard parts (fakes, duplicates, coarse data, explainability) instead of hand-waving past them.

Links

• GitHub repository: github.com/tanishijanweja/SafeSphereAi
• Readme: README.md
• Setup guide: SETUP.md
• Engineering architecture: docs/engineering/architecture.md
• Product : docs/product/preamble.md

Data & API sources:

• data.gov.in — District-wise IPC crime data (NCRB, crimes against women)
• NCRB — National Crime Records Bureau
• GDELT — Global Database of Events, Language, and Tone
• OpenStreetMaps — Maps / Places / Geocoding
• Google Gemini — Gemini API (structured output + embeddings)

Core libraries:

Next.js · Hono · Bun · Prisma · Turborepo · Tailwind CSS · shadcn/ui · Better Auth