AI scam-check + cyber protection for UK families and SMEs
Consumer-grade scam protection — UK's £1.2bn fraud losses gap. AI verdict transparency.
The gap we're addressing.
UK fraud losses topped £1.2bn in 2024, with authorised push-payment fraud, romance scams, smishing and AI-cloned voice impersonation hitting older adults and small businesses hardest. Existing public guidance is reactive and fragmented — victims usually find out after the money has moved. Banks defend the rails, but no consumer-facing AI tool gives an instant, transparent risk verdict on a message, link or screenshot at the moment of decision.
Source: UK Finance, Annual Fraud Report 2024.
What's actually novel.
CyberShield is a multi-modal scam-check engine built on a UK-specific 13-category taxonomy. It accepts text, URLs, screenshots and voice clips, returns a plain-English verdict with reasoning, and is engineered to be readable by an 80-year-old as well as a fraud analyst.
- 13-category UK scam taxonomy (impersonation, smishing, romance, marketplace, AI voice clone, courier, HMRC, investment, parcel, job, tech-support, charity, rental).
- Multi-modal input: paste text, URL, image, or voice — single verdict surface.
- Verdict transparency: every result returns the category, confidence band, and the reasoning chain that produced it.
- Free public tier so the elderly and unbanked are never gated out of fraud protection.
- White-label-ready for councils, charities, credit unions and trading-standards teams.
Honest about early-stage.
Live beta at cybershield.uk with early real usage. Built as the evidence base for cyber-grant applications. The product is live and self-funded; the role of grant capital is to add the rigour, cohort study and independent review that a self-funded team cannot deliver alone.
How the system works.
CyberShield is an LLM verdict engine constrained by a fixed schema. Inputs are normalised to a canonical text representation, then classified against the 13-category taxonomy with structured output. This makes outputs auditable, comparable across versions, and safe to publish.
Ingest
Text, URL, screenshot OCR or voice transcript collapsed into one canonical text + signal payload.
Classify
LLM produces a verdict object: { category, confidence_band, reasoning, red_flags[], advised_action }.
Explain
Reasoning surfaced in plain English with reading-age controls. Action advice is consistent per category.
Log
Anonymised verdicts logged for taxonomy refinement and public scoreboard at /admin/stats.
Three plausible tiers of funding.
Figures illustrative and indexed to UK grant call sizes we're shortlisted against. Each tier delivers a discrete, publishable milestone.
Independent academic review of the 13-category taxonomy, accessibility audit (RNIB / Age UK consultation), and a 200-participant elderly cohort feedback round.
6-month pre-pilot with 2 UK councils and 1 credit union. Comparative accuracy study vs. control. Published methodology paper and v2 taxonomy.
API for banks, councils and trading standards. Voice-clone detection module. Welsh + Urdu + Punjabi reading-age variants. Independent red-team.
Small team. Real product. Open to advisors.
Founder of Sifotech. Operates a UK transport business and three healthtech / cyber consumer products built on the same stack. Solo technical builder; uses our own software in daily operations.
Clinical and cyber advisors welcomed for CyberShield. We're actively looking to formalise the advisory board before the next funding round. We do not list advisors here unless they have signed on — no name-dropping.
Get in touchPrintable evidence pack.
Book a call to discuss the CyberShield round.
15 minutes. We'll walk through this evidence pack live, answer methodology questions, and scope the consortium / advisory fit for your call.