StreetPulse — Urban Safety Intelligence Platform

Predict where West London is risky — street by street, hour by hour — and see it live on a tactical command-center map.

StreetPulse is a full-stack safety platform. A Spring Boot backend fuses community-submitted incident reports with ~6,000 official Police UK crime records (12 months, rate-limit-aware ingestion with caching), runs every new report through an agentic triage pipeline (automatic severity escalation and duplicate merging), and serves it all through a REST API. A Python machine-learning service trains a gradient-boosted spatiotemporal risk model on the fused data — risk that genuinely shifts with the hour of day — and a tool-using AI agent answers safety questions by querying the live data before it speaks. The frontend is a dark, HUD-style command console: radar sweep, target-lock reticles on real incidents, and a scrolling intel feed.

Features

• 🗺️ Live tactical map — Leaflet + CARTO dark tiles with an ML risk heat layer, ~500 police crime markers, community incident pins, a rotating radar sweep, and target-lock reticles that acquire real incidents one by one
• 🔮 ML risk forecasting — a scikit-learn HistGradientBoostingRegressor learns the spatial crime-intensity surface from ~6,000 events, plus a learned hour-of-day effect; scrub a 00:00–23:00 time slider and watch predicted risk shift across the day
• 🤖 Tool-using AI agent — Gemini function-calling with four tools (search_incidents, get_risk_score, get_police_crimes, get_stats); every answer is grounded in live data, with a data-driven fallback so it works with no API key at all
• 🚨 Agentic report triage — reports describing danger ("knife", "followed", "assault"…) are auto-escalated to HIGH severity; near-identical reports (same category, ~150 m, 48 h) are merged into one confirmation instead of cluttering the map
• 👍 Community confirmations — one-tap "Confirm this report" validation with per-device double-confirm protection; most-confirmed reports surface in analytics
• 🔔 Live alerts — recent high-priority activity feeds the notification bell (unseen-count badge), the alerts page, and the scrolling intel ticker
• 📊 Crime intelligence dashboard — KPI cards with animated count-ups, 30-day incident trends, category breakdowns, monthly police-crime trends, area risk index — every card, chart, and row opens a detail popup
• 🔐 Accounts — registration and login with BCrypt-hashed passwords; sessions are browser-scoped so a fresh visit always starts anonymous
• 🛡️ Privacy-aware reporting — anonymous by default; the optional reporter email is write-only (never returned by any endpoint), stored under UK GDPR consent

Architecture

Layer	Tech	Role
frontend/	React 18 · Vite · Leaflet · Chart.js	Command-center UI, live map, analytics, popups
backend/	Spring Boot 3.2 · Java 17	REST API, ingestion, agentic triage, alerts, auth, explainable risk baseline
ml/	Python · FastAPI · scikit-learn	Learned spatiotemporal risk model (gradient boosting + diurnal effect)
database	H2 (file mode)	Incidents, users; self-healing schema maintenance on startup

Police UK API ──► 12-month ingest (429-retry, cached) ──┐
Community reports ──► agentic triage (escalate, dedupe) ─┼──► H2 DB
                                                         │
        Python ML service ◄──────────── trains on ───────┘
        (risk surface + hour-of-day)
              │  /predict · /grid
              ▼
React + Leaflet ◄── REST API ◄── Gemini agent (4 data tools)

The backend degrades gracefully at every layer: no Python service → an explainable kernel-density risk baseline (Gaussian spatial kernel × severity × recency, normalised to a data-derived reference) takes over; no Gemini key → the assistant answers from a data-driven analysis engine. The whole platform runs with zero paid APIs and zero API keys.

REST API

Endpoint	Returns
GET /api/incidents	all community reports (filter by category, severity)
POST /api/incidents	submit a report (runs the triage pipeline)
PATCH /api/incidents/{id}/upvote	confirm a report
GET /api/incidents/analytics	totals, severity split, most-active area, last-24h count
GET /api/police/crimes/recent	~6,000 Police UK records, last 12 months
GET /api/risk?lat&lng&hour	0–100 risk score for a point + hour, with contributing factors
GET /api/risk/grid?hour	scored grid powering the heat layer
POST /api/ai/agent	tool-using agent answer, mode: agent | fallback
GET /api/alerts	recent high-priority activity
POST /api/auth/register · /login	accounts (BCrypt)

ML service (optional, port 8000): GET /health · GET /predict · GET /grid · POST /retrain.

Quick start

Requirements: Java 17, Maven 3.8+, Node 18+. Optional: Python 3.9+ for the ML service.

# Terminal 1 — backend (seeds sample data; police history ingests in background)
cd backend
mvn spring-boot:run

# Terminal 2 — ML service (optional: learned model instead of the baseline)
cd ml
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
uvicorn app:app --port 8000

# Terminal 3 — frontend
cd frontend
npm install
npm run dev

Open http://localhost:5173. Everything works immediately with no keys. To enable the live AI agent, export a free GEMINI_API_KEY (https://aistudio.google.com/apikey) before starting the backend.

Data sources

• Police UK Open Data — street-level crime records (Open Government Licence v3.0)
• Community reports — submitted in-app, triaged automatically
• OpenStreetMap / CARTO — dark basemap tiles
• Google Gemini (free tier, optional) — agent reasoning

Privacy

• Anonymous reporting by default; optional email is write-only and never exposed by the API
• Coordinates stored at street-level precision; no tracking, no cookies
• Sign-in is session-scoped — a fresh visit always starts with no personal details