## Inspiration Seniors lose over $28 billion a year to financial fraud, and most of it happens silently. We were inspired by real stories of grandparents getting tricked by gift card scams, fake IRS calls, and phishing emails — and family members who had no idea until it was too late. We wanted to build the safety net that didn't exist.
## What it does SafePay Family is an AI-powered financial protection platform for seniors. It scores every transaction in real time using machine learning, instantly alerts trusted family members when something looks suspicious, and lets anyone check a sketchy email, text, or phone number through our Fraud Detection tool — returning a SCAM or SAFE verdict with a confidence score and breakdown of warning signals.
## How we built it We built a full-stack app with three layers. The frontend is Next.js + TypeScript with Tailwind, with separate dashboards for seniors and family members. The backend is Node.js + Express with Prisma and PostgreSQL (hosted on Neon) handling auth, alerts, and transactions. The ML service is a Python FastAPI app running two models — an Isolation Forest for anomaly detection and a Logistic Regression for fraud probability — combined into a single risk score. The scam detector uses TF-IDF + Logistic Regression trained on phishing and spam patterns. Real-time alerts are powered by Socket.io, and AI explanations use the Gemini API.
## Challenges we ran into Getting the ML pipeline to work end-to-end in real time was the hardest part — engineering features from raw transactions fast enough to not block the payment flow. We also struggled with class imbalance in the fraud dataset and tuning the anomaly score normalization so it wasn't too sensitive or too lenient. On the frontend, syncing real-time socket alerts across both dashboards without stale state was trickier than expected.
## Accomplishments that we're proud of We're proud that the whole system actually works end-to-end — a transaction submitted on the senior dashboard gets ML-scored, triggers a real-time alert on the family dashboard, and gets an AI-generated explanation, all in under two seconds. We're also proud of the scam detector being genuinely useful — it catches typosquatting, government impersonation, urgency language, and suspicious domains out of the box.
## What we learned We learned how to combine supervised and unsupervised ML into a single interpretable risk score, and how important feature engineering is compared to model choice — the behavioral baseline features (amount ratio, velocity, new merchant flags) did more work than the model itself. We also learned how much thought goes into designing for seniors specifically — simplicity, clear language, and not overwhelming them with alerts matters a lot.
## What's next for SafePay Family We want to add phone call screening using voice AI to catch live scam calls in real time. We also want to integrate directly with real banking APIs so protection is automatic rather than transaction-by-transaction. On the ML side, we'd retrain on a larger real-world fraud dataset and add a feedback loop so family decisions improve the model over time. Longer term, a mobile app with push notifications would make the family alert system much more responsive.
Log in or sign up for Devpost to join the conversation.