About SheriaLens

Kenya's first AI-powered civic intelligence platform. We verify political claims against Kenyan law, detect misinformation networks, monitor civic discourse, and deliver the truth — in English, Kiswahili, and Sheng — across every channel Kenyans use.

Our Mission

In 2024, misinterpretation of the Finance Bill provisions spread faster than corrections could reach people — contributing to nationwide protests, parliament breach, and loss of life. SheriaLens exists to ensure that never happens again. We give every Kenyan — from university students to farmers, from WhatsApp users to X followers — instant access to verified civic information before misinformation can take hold.

What We Do

📋

Fact-Check Claims

Submit any political claim via X, WhatsApp, Telegram, or our web portal. Our AI verifies it against 275,000+ Kenya Law documents and returns a verdict with citations.

⚖️

Legal RAG Search

Natural language search across the Constitution, Acts of Parliament, Bills, and judicial decisions. Ask questions in English or Kiswahili — get answers with specific section references.

🚨

Early Warning System

Real-time crisis pattern detection comparing current digital activity against fingerprints from Kenya's 5 historical crises (2007, 2013, 2017, 2022, 2024). Alerts government analysts before situations escalate.

🕸️

Network Intelligence

OSINT analysis tracking who spreads misinformation, how far it reaches, and whether coordinated inauthentic behavior (bots) is involved. Identifies superspreaders and generates counter-messaging.

🔍

Deepfake Detection

AI-powered analysis of images and media for synthetic manipulation, recycled imagery, and doctored documents. EXIF metadata extraction and reverse image search.

🎓

Civic Education

Bill tracker showing legislation through Parliament, personalized impact analysis for your profile (student, farmer, SME owner), public participation guides, and civic quizzes.

Verification Pipeline

Every claim follows this rigorous 10-step verification process, aligned with the International Fact-Checking Network (IFCN) Code of Principles.

1

Claim Reception

Claims arrive via X (@SheriaLens), WhatsApp, Telegram, web portal, or API. All channels normalize to a unified ClaimRequest format.

2

Language Detection

Automatic detection of English, Kiswahili, Sheng, or code-switched text. Responses are always in the same language as the submission.

3

Claim Extraction & NER

Named Entity Recognition identifies politicians, bills, organizations, monetary amounts, and dates. Compound claims are decomposed into individual verifiable assertions.

4

Check-Worthiness Scoring

AI classifier scores how "checkable" the claim is (0.0-1.0). Opinions, predictions, and value judgments are identified as non-fact-checkable and explained to the user.

5

Deduplication

SHA-256 hash matching against previously verified claims. If the same claim was already fact-checked, the existing verdict is returned instantly (>0.92 cosine similarity).

6

Evidence Retrieval (Parallel)

Simultaneous search across: Kenya Law corpus (275K+ documents via Legal RAG with hybrid BM25+vector search), Google Fact Check Tools API, GDELT news archive, and government data sources.

7

LLM Verification

Claude Sonnet performs legal reasoning: cross-references the claim against retrieved evidence, determines verdict, generates confidence score with transparent breakdown (source reliability × evidence strength × claim complexity).

8

Human Review Routing

Claims involving sitting presidents, governors, ethnic dimensions, ongoing court cases, or elections are flagged for human expert review. Three-reviewer consensus required for politically sensitive claims.

9

Verdict Assignment

One of five verdicts: TRUE, FALSE, MISLEADING, MIXED, or UNVERIFIED. Each includes specific legal citations, confidence score, and plain-language explanation.

10

Multi-Channel Delivery

Verdict formatted for the originating channel (280 chars for X thread, rich message for WhatsApp, markdown for Telegram, full page for web) with citations and share links.

Verdict System

TRUE

Claim fully supported by authoritative legal sources. Specific sections cited.

FALSE

Claim contradicted by authoritative sources, OR the claimed provision does not exist in the referenced legislation.

MISLEADING

Technically true but presented with deceptive context that changes the meaning.

MIXED

Contains both accurate and inaccurate elements. Each part is addressed separately.

? UNVERIFIED

Claim is about something genuinely unknowable from available sources (future predictions, private intentions). Only used when no relevant evidence exists.

Early Warning System

SheriaLens doesn't just react to misinformation — we predict it. Our early warning system compares current digital activity patterns against fingerprints from Kenya's historical crises:

2007Post-Election Violence

Ethnic incitement via SMS/radio, rapid geographic spread

2013Cambridge Analytica

Foreign-backed social media manipulation, coordinated fake accounts

2017Election Disinformation

87% of Kenyans encountered disinfo, multi-platform coordination

2022Influencer Industry

Domestic paid influencer campaigns, county-level targeting

2024Finance Bill Protests

Gen Z mobilization, viral bill misinterpretation, rapid velocity

Using cosine similarity matching across 6 dimensions (sentiment velocity, volume anomaly, negative ratio, geographic concentration, CIB signal, mention spike), the system classifies current activity as GREEN (normal), YELLOW (elevated), ORANGE (high), or RED (crisis match) — and generates counter-narrative recommendations for each threat level.

Technology Stack

Legal RAG

275,000+ Kenya Law documents. Hybrid BM25 + dense vector search via Weaviate. RCTS chunking respecting legal hierarchy.

LLM Reasoning

Claude Sonnet for legal analysis and verification. Politically neutral system prompts. Prompt caching for 80% cost reduction.

Sentiment Analysis

AfriBERTa for Kiswahili, AfroXLMR for code-switched text. Real-time scoring of social media posts across X, Telegram, and news.

Network Analysis

Temporal clustering, content duplication detection, and bot scoring to identify coordinated inauthentic behavior.

Multi-Channel

X bot (@SheriaLens), WhatsApp, Telegram (@SheriaLensBot), web portal — all sharing the same verification pipeline.

Infrastructure

Event-driven microservices on Kafka. PostgreSQL + Weaviate + Redis. 19 services, 4 bots, 4 web apps.

Political Neutrality

Political neutrality is enforced in code, not just policy. Our LLM system prompts explicitly prohibit opinion generation, ethnic framing, and partisan language.

  • We never say "the government claims" (implies doubt) — we cite the source directly
  • We present both government and civil society perspectives where applicable
  • We never frame issues along ethnic or regional lines
  • Every AI response is logged for quarterly external bias audits
  • Public participation templates guide structure, never pre-fill positions

Corrections Policy

If SheriaLens publishes an incorrect verdict, we issue an immediate public correction on the same channel. Corrections are never deleted — the original verdict is preserved with a "CORRECTED" label. All corrections are logged in our public corrections register.

Privacy & Data Protection

SheriaLens is compliant with Kenya's Data Protection Act 2019 (DPA 2019). No real names are stored. Phone numbers are SHA-256 hashed. Users can DELETE MY DATA or EXPORT MY DATA at any time from any channel. Data retention is 12 months maximum with automated purging.

Read our full Privacy Policy →

Standards & Partners

SheriaLens follows the IFCN Code of Principles and studies patterns from PesaCheck, Africa Check, Meedan Check, and UNDP iVerify. Every published fact-check generates a schema.org/ClaimReview for Google and Facebook indexing.

Have a claim to verify?

Submit it on any platform — we'll check it against Kenyan law and get back to you.