Engines – Alfred Kinoti

VMIP · Nairobi Estate Intelligence · Metahac Labs

VMIP NAIROBI Estate Intelligence

Estates

Sub-Cty

Incidents

Critical

LIVE

--:--:--

Map View

Filter by Sub-County

Estates by Risk

Map Layers

Heatmap

Sub-County Zones

Estate Markers

Anomaly Alerts

Community Reports

Coverage Gaps

Hawkes Risk Grid

Estate Labels

Base Map

Risk Classification

Critical

>0.82

High

0.68–0.82

Medium

0.55–0.68

Low

<0.55

Estate-level figuresSynthetic baseline

City/county figuresCited, real

View real data sources \u2192

Metahac Labs · Alfred Kinoti

Nairobi › ›

Mini-map

Report Incident

Incident Log

Coverage Gaps

Real Data Sources

Ethical Governance

ACLED Live Data

Generate Report

Advanced Analytics

Estate

—

Risk Score

—

Incidents

—

Click to open intelligence panel

VMIP-2025 · 72 Estates · 17 Sub-Counties · Metahac Labs · Alfred Kinoti

— › Estate Level

SELECT AN ESTATE

— RISK — — LIVE INTEL

Summary

—

Incidents

—

Risk/100

—

Top Crime

Risk Score

—

/ 100

——

LOWMEDHIGHCRIT

Crime Breakdown

Incident Trend

Time of Day

Confirmed Reports

Community Reports

Human-submitted, not AI-generated. Unverified reports are visible but excluded from the estate's official incident count and risk score until verified by a reviewer.

Community-Verified Hotspots

Built entirely from your own Verified community reports, grouped by location. This is the only genuinely real, current source of alley/footpath/corridor-level detail \u2014 news coverage almost never names these.

Documented Hotspots (research/news)

Modus Operandi

Police Response

Intelligence Assessment

Profile Tags

AI Explainability

Decision-support output only. Not grounds for arrest, stop, or search on its own. A human reviewer must validate before any action is taken.

No explanation generated yet. Click "Generate Explanation" to query contributing factors via the Anthropic API (SHAP-style / LIME-style / counterfactual).

Public Sentiment

Aggregated sentiment from public news/community sources only. Used for resourcing and community-engagement context — never for identifying or targeting individuals, and never for monitoring lawful assembly or political speech.

No sentiment analysis run yet. Click "Analyze Sentiment" to pull public news/community signal via the Anthropic API.

NLP Model Reference

Nearby Estates

Live Intelligence Scrape

Uses Claude + web search to pull latest news and police reports for this estate.

Anthropic API Key

Key stays in-browser only, reused for this session · Get free key

Searching sources…

Ethical Governance Layer

VMIP is a decision-support tool. These constraints are enforced by policy and apply to every prediction, explanation, and sentiment output in this app.

This system never

This system is designed to

Report an Incident

Real, human-submitted data. Stored locally in this browser only. New reports start as Unverified and don't affect risk scores until a reviewer verifies them.

Estate

Incident Type

Date / Time

Severity

Low

Medium

High

Critical

Location / Landmark (optional)

This is the most reliable way to get real alley/footpath/corridor-level detail \u2014 published news almost never names these, only people on the ground do.

Description

Your Name / Contact (optional)

Incident Log

All community-submitted reports across every estate, stored locally in this browser.

All

Unverified

Verified

Coverage Gap Analysis

Enter your actual current patrol/guard allocation per estate (units, hours/week, whatever you track). The app weighs that against modeled risk and population density to surface where coverage doesn't match need. Decision-support only \u2014 final resourcing is a human call.

Estate	Risk	Density	Allocated	Needed	Gap

Gap \u2265 3 \u2014 under-resourced Gap 1\u20132.9 \u2014 mild gap Gap \u22121 to 1 \u2014 balanced Gap < \u22121 \u2014 over-resourced

Real Data Sources

No public source breaks Nairobi crime down by individual estate \u2014 that granularity doesn't exist. What's below is real, citable data at the levels that are published: national, county, and police division. The 72-estate map elsewhere in this app uses illustrative Synthetic baseline figures, sharpened only by your Verified community reports and SCRAPE-sourced news.

Nairobi City crime volumeCounty

9,717

Crimes recorded, 2024

11,108

Crimes recorded, 2023

9.6%

Share of national total

Nairobi City Police Command recorded the highest crime count of any county in 2024, though down from 2023.

Source: KNBS Economic Survey 2025

National crime trendNational

96,038

Total cases, 2025 (latest)

101,220

Total cases, 2024

National recorded crime has trended downward year over year. These are total Kenya figures, not Nairobi-specific.

Source: KNBS Economic Survey

Victimization survey, by crime typeCounty

Share of surveyed Nairobi County households reporting having experienced or perceived each crime type \u2014 not incident counts.

Source: National Crime Research Centre (NCRC) victimization survey

Police division-level trendDivision

A private security-sector annual report (not raw police records) found crime was most prevalent in the Buruburu Division in December 2024, with Kasarani, Langata, Makadara, Ruai and Kayole Divisions showing a steady rise through Q4 2024. This is the closest publicly available data gets to sub-county granularity for Nairobi \u2014 still far coarser than individual estates.

Source: RFSSL Annual Crime Report 2024

Figures are illustrative of what's publicly documented at time of writing and may shift as newer reports are published \u2014 verify against the original source before using for operational decisions.

ACLED Live Data \u2014 Nairobi

Fetches real political violence, riot and demonstration events for Nairobi County from ACLED. This is NOT street crime data \u2014 ACLED covers armed conflict, political violence, and protest, not individual robbery or theft. Requires a free myACLED account (email + API key).

Connecting to ACLED\u2026

Generate Intelligence Report

Produces a downloadable HTML report (open in browser and print to PDF). Synthetic baseline figures are clearly marked.

Report Sections

Estate overview & risk scores

Intelligence dossiers (real, cited)

Hotspots (research + community-verified)

Community verified reports

Coverage gap analysis

Cached AI results (Explain/Sentiment/Scrape)

Citywide real statistics

Estates to Include

Report Details

Prepared by

Classification / Handling

Notes / Executive Summary

Advanced Analytics Engine

Mathematical crime prediction, network mapping, and patrol optimization. All outputs are decision-support — human review required before any operational action.

λ Hawkes Process

R-GCN Network

Stackelberg Patrol

📡 Live Data

Core Master Equation — Conditional Intensity Function

λ(t,x,y) = μ(x,y) + ∑_{i: tᵢ<t} g(t−tᵢ, x−xᵢ, y−yᵢ) g(Δt,Δx,Δy) = θ · ω · e^−ωΔt · (1/2πσ²) · e^{−(Δx²+Δy²)/2σ²}

λ(t,x,y) conditional risk intensity at location (x,y) at time t · μ(x,y) static background rate (IDW from estate risk scores) · θ productivity — expected after-crimes per event · ω temporal decay rate (hr⁻¹) · σ spatial scale (km). Fitted by EM (Veen & Schoenberg branching-structure algorithm), not gradient descent — the standard, numerically stable method for Hawkes MLE.

Property Crime

Violent Crime

Drug Crime

Parameters (θ, ω, σ) — adjust manually or click Fit Model to estimate via EM on a 70/30 train/test split

θ Productivity0.40

ω Temporal decay0.08 hr⁻¹

σ Spatial scale0.80 km

—

Log-Likelihood L

—

AIC (3 params)

—

Branching ratio n*

—

PAI@10% (held-out)

Log-Likelihood (EM Objective) & Stability Condition

L = ∑_i ln λ(tᵢ,xᵢ,yᵢ) − ∫₀ᵀ∫∫Ω λ(t,x,y) dx dy dt n* = θ (branching ratio — process is stable iff n* < 1, explosive otherwise)

PAI = (n/N) / (a/A), evaluated only on the held-out 30% test window — fitting and scoring on the same events would be data leakage. PAI > 1 beats random; PAI > 5 indicates strong spatial concentration.

Calibration — Predicted vs. Observed Share, by Risk Decile

If the model is well-calibrated, predicted risk share (blue) and the held-out events actually observed in that decile (green/amber/red) should roughly match for every decile.

Top Predicted Intensity — λ(T, x, y) at grid cell centres

Estate / Cell	μ(x,y) Base	Excitation	λ Intensity	Intensity Bar

R-GCN Layer Update — Relational Graph Convolutional Network

hᵢ⁽ˡ⁺¹⁾ = σ( W₀ˡ hᵢˡ + ∑_r∈R ∑_j∈Νᵢʳ (1/cᵢ₃) Wᵣˡ h⸉ˡ )

Νᵢʳ neighbours of node i under relation r · Wᵣˡ relation-specific transformation matrix · cᵢ₃ normalization = |Νᵢʳ|. Edges encode two real relation types \u2014 geographic proximity (<2.5 km) and shared sub-county \u2014 with a third "both" channel for pairs where both hold. σ = ReLU. Node features h₀ = [risk score, log-scaled incident count, sub-county avg risk], all real data, not random.

Honest disclosure Weight matrices W are fixed/hand-specified, not learned by backpropagation \u2014 there is no labeled training signal (e.g. confirmed-syndicate-link ground truth) available to train R-GCN weights against. This runs the real forward-pass equation over real graph structure and real features; it does not claim to be a trained classifier.

Click any node to open that estate's dossier. Node size ∝ baseline risk score. Dashed red ring = network-adjusted risk exceeds baseline after 2-layer message passing.

Sub-county + geo Same sub-county Geo proximity only

Top Network-Contagion Effects (2-layer R-GCN output)

Stackelberg Security Game — Mixed Strategy Computation

max_x ∑_p∈P xᵣ · U_d(p, a*(x)) s.t. ∑_p xᵣ = 1, xᵣ ≥ 0 a*(x) = arg max_a∈A ∑_p xᵣ · U_a(p,a)

xᵣ mixed-strategy probability for patrol route p · U_d(p,a) = +V(a) if p covers a, else −V(a) · V(a) = risk(a)×density(a). Solved via Multiplicative Weights Update (Freund & Schapire 1997) \u2014 a no-regret algorithm whose time-averaged strategy is guaranteed by the minimax theorem to converge to the zero-sum game value, avoiding the need for a full MILP. Σxᵣ = 1.0 is verified and displayed after each solve. Random sampling from this mixed strategy produces patrol schedules that are provably unpredictable to a rational adversary.

Available patrols

#	Estate	Sub-Cty	xᵣ probability	%

What this is and isn't: GDELT monitors ~100,000 global news sources every 15 minutes and covers Kenya violence/crime events. ACLED documents political violence and unrest with real coordinates. Both are real, open, citable data — but they cover news-reported incidents and political violence, not granular street crime (robbery, theft at estate level), which has no open data source for Nairobi. Claude classifies and geocodes incoming articles to the nearest estate. Events with confidence ≥ 60% are merged into the Hawkes training pool alongside your verified community reports.

New events ready

GDELT

15-min global news
Free · No auth · CORS-ok
Last: never

ACLED

Political violence
Free · myACLED key
Last: never

Community

Human-submitted
Verified only · Local
All reports

Auto-poll:

min confidence %

Training Pool Composition

Recent Live Events