Designing Trust in AI: A Scam Detection Product
Project Overview
Vigil is an AI-powered scam detection app designed to help users evaluate suspicious messages confidently.
As the Founding Product Designer, I led the design of a 0→1 MVP that translates a probabilistic AI risk model into a clear, user-facing decision system.
The product allows users to submit messages via paste or email, receive an immediate risk assessment, and optionally refine results with additional context before receiving actionable guidance.
The MVP was developed for pre-seed investor presentations, with a focus on demonstrating the product’s core functionality and value.
Problem
Online scams are one of the most aggressive forms of fraud, targeting everyday communication forms like texts and emails.
Obviously scams are designed to look legitimate, mimicking trusted brands, account alerts, and even friends or family.
For most people, identifying these scams can be pretty difficult. Fraudulent messages often rely on subtle cues such as suspicious links, social mimicry tactics, or urgent language that can be hard to recognize in the moment.
As a result, people often face two equally problematic outcomes:
• False trust: assuming a fraudulent message is legitimate
• False alarm: panicking about a message that is actually safe
Vigil developed a patented AI model that analyzes messages and predicts scam likelihood using a risk score from 0–100 and a confidence classification.
The project challenge wasn’t detection but translation. Most users aren’t familiar with probabilistic risk systems, making raw AI outputs easy to misinterpret.
My role was to design an experience that transformed complex AI analysis into clear, calm decision support users could trust.
Product Challenge
While Vigil’s AI model could analyze messages and generate a risk score, translating that output into a usable product experience introduced several challenges.
Most users are unfamiliar with probabilistic risk systems, making numerical scores easy to misinterpret. A risky message could trigger panic, while moderate results might be ignored or misunderstood.
Additionally, message content alone was not always enough for accurate analysis. Certain signals such as whether a sender was expected or recognized required user input, introducing a second challenge: how to gather meaningful context without adding friction.
The product needed to balance:
• Speed: delivering immediate insight with minimal effort
• Accuracy: allowing deeper analysis when needed
• Clarity: making results easy to understand and act on
The challenge was designing an experience that supported both quick decisions and deeper investigation without overwhelming users.
My Role
As Vigil’s founding product designer, I was responsible for translating a complex AI scam detection model into a clear, intuitive product experience.
Working closely with the founder/CEO and engineering, I led the end-to-end product design for the MVP defining information architecture, user flows, onboarding, risk interpretation, and the product’s multi-stage interaction model. I also created wireframes, prototypes, high-fidelity designs, a scalable design system, and collaborated on branding and technical implementation.
Because Vigil was a pre-seed MVP designed for investor presentations, decisions prioritized clarity, feasibility, and communicating value quickly. Due to NDA constraints and limited user testing, design decisions were guided by UX heuristics, risk perception psychology, and close collaboration with the founding team.
Information Architecture
A three-stage product system designed to balance speed, clarity, and depth
Sentinel: Initial Assessment
Instant AI analysis provides a risk score (0–100) with a clear visual and textual interpretation, prioritizing speed and clarity.
Scout: Contextual Refinement
An optional refinement layer that gathers user-provided context to improve accuracy without adding unnecessary friction.
Sage: Decision Support
Clear, actionable recommendations that help users confidently decide how to proceed.
Key Product Decisions
Designing Risk as a Spectrum
Rather than presenting scams as simply “safe” or “unsafe,” I designed risk as a continuous spectrum (0–100) to reflect the probabilistic nature of the AI model. This reduced false certainty and gave users a more nuanced understanding of risk.
Making Risk Interpretable at a Glance
A numerical score alone creates ambiguity, especially in high-stakes situations. To improve clarity, I paired the score with a gauge visualization, color-coded spectrum, and interpretation labels, making risk easier to understand at a glance while reinforcing that outcomes could shift with additional context.
Balancing Speed with Accuracy
Because the model improved with user-provided context, I designed Scout as an optional refinement layer rather than a required step. This allowed users to receive immediate insight while supporting deeper analysis when uncertainty remained.
Separating Analysis from Action
I split the experience into Sentinel (analysis) and Sage (guidance) to reduce overwhelm and support clearer decision-making. Users first understand the system’s assessment before receiving actionable recommendations.
System Design
A scalable design system built to support consistency, clarity, and future growth
To support Vigil’s multi-stage experience, I designed a scalable system of reusable components, interaction patterns, and risk states tailored to the needs of the MVP
Component System
I created reusable modules for risk scoring, recommendations, inputs, and interactions, ensuring consistency across Sentinel, Scout, and Sage while enabling faster iteration and engineering collaboration.
Risk Communication
Because the product centered around interpreting uncertainty, I developed a risk-based color and scoring system that remained consistent across all stages. Shared visual cues like the score, label, and color, reduced cognitive load and made results easier to interpret at a glance.
Designed for Growth
The system was intentionally built to scale, supporting future refinement flows, expanded recommendation logic, and new communication channels without requiring major structural changes.
Next Steps
Future validation focused on trust, interpretation, and decision-making under uncertainty.
Because Vigil was developed as a pre-seed MVP under NDA, formal testing and product metrics were limited. If moved into a live environment, validation would focus on three key areas:
Risk Interpretation
Testing how users understand mid-range scores, interpretation labels, and visual signals to ensure risk is communicated clearly in uncertain scenarios.
Behavior Across the System
Evaluating when users choose to engage with Scout’s optional refinement layer and whether additional guidance improves decision-making in higher-risk situations.
Trust and Outcomes
Measuring how users respond to recommendations ensuring the system supports confident decision making without creating panic or overconfidence.
Key takeaways
Designing AI products means designing for uncertainty, trust, and human behavior not just for usability.
Designing for Uncertainty
Working with probabilistic AI reinforced that product design is not always about right or wrong answers. It’s about helping users make confident decisions in uncertain situations.
Designing Systems, Not Screens
The challenge was not designing individual interfaces, but translating a complex AI model into a structured product experience with clear flows, interpretable outputs, and actionable guidance.
Balancing Tradeoffs
Vigil reinforced that strong product design often means balancing competing needs like speed with accuracy, simplicity with depth, and automation with human judgment.