From Raw Report to Safety Score: Inside Chipon's Intelligence Pipeline

When you open Chipon and see a safety score of 74 for your neighborhood, that number represents the output of a process that ingests data from five distinct sources, weighs them against each other, applies freshness decay, and computes a score — all within minutes of new data arriving.

Today, we're pulling back the curtain on how it works.

The Five Data Sources

Chipon's safety intelligence doesn't come from a single stream. It's a composite of:

1. Community Reports — Real-time incident reports filed by Chipon users. These are our most valuable data source because of their immediacy and geographic precision.
2. News Intelligence — We process articles from over 40 Nigerian news outlets using natural language processing to extract incident data, locations, and severity classifications.
3. ACLED Data — The Armed Conflict Location & Event Data Project provides conflict event data that we pull weekly. This captures political violence, protests, and security force incidents.
4. HDX Humanitarian Data — From the Humanitarian Data Exchange, we ingest displacement, flooding, and infrastructure failure data.
5. NBS Baseline — The Nigerian Bureau of Statistics crime data serves as our historical baseline, calibrating expected incident rates by region.

The Ingestion Pipeline

Each data source enters Chipon through its own ingestion adapter. Here's the simplified flow:

Source → Adapter → Normalization → Geolocation → Deduplication → Classification → Storage → Score Computation

The critical steps are geolocation and deduplication. A news article about a robbery in “Surulere” needs to be mapped to a precise geographic point, not just a neighborhood name. And when three news outlets cover the same incident, we need to recognize it as one event, not three.

AI-Assisted Classification

For news-sourced incidents, we use AI to classify the category (armed robbery, accident, protest, etc.) and estimate severity. The model assigns a confidence score from 0-100. Incidents above 70% confidence are auto-classified. Those below go into a review queue.

Community reports, by contrast, are classified by the reporter themselves. But we still run them through a sanity check — if someone reports a “fire” with a description that sounds like a traffic accident, we flag it for verification.

The Scoring Algorithm

The neighborhood safety score is computed fresh every 24 hours (and incrementally with every new community report). Here's the formula in plain English:

1. Start at 100 (perfect safety).
2. Subtract penalties for each incident in the area over the past 90 days.
3. Weight by severity: a critical incident (armed robbery in progress) penalizes 4x more than a low-severity one (road closure).
4. Weight by freshness: an incident from today penalizes more than one from 60 days ago. We use linear decay over the 90-day window.
5. Weight by source reliability: community-verified incidents (3+ confirmations) count 1.2x. ACLED data counts 1.5x. Unverified reports count 0.7x.
6. Compute sub-scores for crime rate, lighting, police proximity, and community engagement.
7. Cap the final score between 0 and 100.

Why This Matters

The difference between Chipon and a simple crime map is intelligence. A crime map shows you dots. Chipon shows you a score that synthesizes multiple data streams, weighs recency and reliability, and gives you a single number you can act on.

When you check Chipon before heading out, you're not looking at a dot on a map. You're consulting a system that has processed thousands of data points to give you the most accurate picture of what's happening around you right now.

Interested in the technical details? We'll be publishing more behind-the-data articles diving deeper into individual components of the pipeline. Follow the Chipon blog to stay updated.