Guardian Labs Stops Scam Calls Before the Phone Rings

Rita Barbosa watched her grandmother lose roughly €4,000 to a phone scam that sold four water filters worth about €20 each for €1,000 apiece. The shock was not only the money. Her grandmother stopped answering the phone and lost confidence in technology. That reaction became the seed for Guardião, the on-device AI layer from Guardian Labs that now aims to silence fraudulent calls and messages before a Portuguese phone even rings.

Less than a year after a 48-hour hackathon win in Hamburg, the three-person startup holds a formal protocol with the Portuguese Public Security Police, research ties to leading engineering universities, accelerator support, pre-seed funding, and a public endorsement from President António José Seguro. The consumer app stays free. Banks, insurers and telecoms are expected to pay for the campaign intelligence that sits one layer upstream.

The Call That Never Rings

Guardião sits as an AI filter between an incoming call or SMS and the user. Models run entirely on the device. Nothing leaves the phone for remote analysis. For messages the system inspects content before any notification appears, scanning for spoofing signals, phishing links and manipulation patterns. For voice it acts earlier still.

Known scam data is checked first. Content is then scored for artificial urgency, emotional pressure or requests for banking codes. If the risk crosses the threshold the call never rings and the message never surfaces in the normal inbox. A quiet notification lands instead. Inside the app users see a full history of blocked contacts, the stated reason and a risk level.

• Spoofed sender identity and known scam numbers
• Phishing or malicious links inside SMS
• Language that creates false urgency or family crisis
• Requests for one-time banking codes or remote access
• Patterns drawn from police reports and user feedback

Barbosa describes the effect as proactive and invisible. Install once, grant permissions, and the layer keeps running. Similar on-device approaches have appeared in big platforms, including Google’s on-device scam detection features and WhatsApp’s on-device scam alert system. Guardião’s bet is full local control plus Portuguese scam texture.

Forty-Eight Hours in Hamburg

Barbosa holds a computer science and engineering degree from NOVA FCT and was studying artificial intelligence at Hamburg University of Technology when she entered the AI BEAVERS hackathon, also described as the Cursor 2-Day AI Hackathon. Roughly 400 to 450 participants from 40 to 48 countries had 48 hours. Her team’s Guardião prototype took first place and the social voting award. Prizes totalled about $70,000 in cash, credits and support packages.

The personal driver was direct. Scammers convinced her grandmother the household water was unsafe and that she was contaminating her family. Four expensive filters followed. Barbosa later told interviewers the deeper wound was independence: her grandmother decided technology was no longer safe to handle alone.

1. Early 2026: Grandmother loses roughly €4,000 in the water-filter telephone scam.
2. February 2026: Guardião prototype wins the Hamburg AI hackathon in 48 hours.
3. Spring 2026: Viral growth in Portugal reaches nearly 20,000 Instagram followers; pre-seed raised; PSP protocol signed.
4. Mid 2026: President of the Republic publicly endorses the mission; international bank and telecom pilots move toward close.

NOVA FCT later highlighted that Barbosa won first place at Europe’s largest AI hackathon with an app built to protect older users who struggle to separate legitimate calls from fraud. The technical report noted more than two million people over 65 in Portugal.

Portugal’s Favourite Scams Travel Badly

Tactics shift by country. Barbosa stresses that models trained only on one market miss the local flavour that makes a message feel plausible. Two patterns dominate Portuguese reports.

Barbosa estimates the healthcare variant alone funnelled around €2 million to €3 million into mule accounts in a recent year. Cross-border intelligence lets the company spot a campaign migrating from Germany into Portugal or the reverse and retune detection before volume peaks. PSP data from earlier years already showed scams costing more than €110 million in asset losses in 2023 alone, a sharp rise from the prior year.

Who Pays So Families Pay Nothing

Guardian Labs rejects a paid consumer model. Barbosa calls this layer of protection “like a human right.” Too much AI already creates fear and friction; charging families to stay safe feels wrong. Revenue is designed to come from B2B contracts with banks, insurers and telecom operators. Those institutions already absorb refund costs, insurance payouts and legal fallout.

The second product is an intelligence platform. It surfaces active international scam campaigns so partners can flag mule accounts, the temporary bank accounts used to receive and move stolen money. Freezing those accounts before victims transfer funds multiplies the effect of every individual block. The company is closing its first international pilots with banks and telecoms. Public-authority relationships stay partnerships rather than paid contracts, preserving the free consumer path.

Portuguese product studio Subvisual has joined as development partner, investor and venture partner, adding engineering depth to the three-person core team. CTO is Barbosa’s brother Tiago, a prior startup founder with cybersecurity background. She has known him her entire 25 years.

When the Model Is Wrong

False positives exist. Barbosa’s priority at this stage is avoiding false negatives, the scams that slip through. She frames a 99 percent accuracy claim with continuous retraining on police data plus user reports. The remaining 1 percent is visible: every blocked item appears in the dashboard with its reason. Users can recover a legitimate message or call without permanent loss of data.

• Priority: Block real scams first; false negatives cost money and trust.
• Mitigation: Continuous training on PSP reports and in-app feedback.
• User recovery: Full history and risk labels remain accessible inside the app.
• Claimed accuracy: High enough that the residual false positives feel manageable.

Someone who falls for a scam loses far more than someone who later reopens a blocked message. That trade-off is explicit.

Police Data Meets a Three-Person Bench

The Portuguese Public Security Police supplied large volumes of historical scam reports for model fine-tuning. The same force has already name-checked the app in international settings. At a Global Anti Scam Alliance conference in Lisbon the PSP highlighted partnerships including Guardião while discussing transnational burlas and spoofing with Europol and U.S. counterparts.

Two leading Portuguese engineering universities supply research cooperation. One of the country’s largest accelerators provides additional backing. The presidential endorsement arrived as a public “Guardian” gesture: President António José Seguro reviewed the proactive, invisible blocking approach and associated himself with the mission. Growth after the February win was rapid. Nearly 20,000 Instagram followers arrived in months because almost every Portuguese conversation about the app surfaces a cousin, parent or sibling who has already been hit.

Freezing the Mule Account Upstream

Individual blocks protect one household. Campaign intelligence protects the next thousand. When Guardião and partner feeds identify a repeating healthcare SMS pattern or a fresh “lost phone” wave, the institutional platform can push alerts that let banks freeze receiving accounts before the bulk of victims transfer funds. That is the second-order move: stop the money trail rather than only the ringtone.

This type of protection is really needed. It’s like a human right. We already have so much AI controlling and scaring us, creating so many different types of problems, that having to pay for this kind of protection just doesn’t sit right with us. It should be free.

Barbosa made the point in the Tech.eu interview that introduced the company to a wider European audience. The same logic underpins future ambitions to embed the local models inside messaging platforms so a scam that hops from SMS to WhatsApp to Instagram stays visible to a single on-device guardian.

Global enforcement is already scaling. Operation First Light 2026, coordinated by INTERPOL across 97 countries, produced 5,811 arrests and USD 293 million intercepted. That same operation is covered in depth as Interpol’s record global fraud crackdown. Local tools that feed those networks with fresher, privacy-preserving signals close the loop between the grandmother who no longer answers and the mule account that never receives the next transfer.

Guardian Labs still has pilots to finish and an accuracy curve to prove at scale. The architecture is already clear: keep the defence invisible and free on the phone, sell the pattern intelligence to the institutions that currently write the refund cheques, and give older users their independence back one silent block at a time.