FORGEHACKS ONLINE 2026 · AI + CYBERSECURITY
Follow the evidence.
A working demo of ScamTrace: connect suspicious scholarship offers, inspect the sources, and challenge mistakes.
From a suspicious offer to inspectable evidence.
See shared payment destinations, human corrections that retract unsupported connections, screenshot OCR with individual review, local MiniLM narrative retrieval, synthetic evaluation, and portable case exports.
All offers and identifiers in this demo are fictional. Connections are investigative leads; they are not fraud verdicts. Evaluation results describe synthetic data, not real-world fraud accuracy. Narration is synthesized.
Read the demo transcript
Follow the evidence
Scholarship scams can change their names, websites, and wording. What often stays familiar is the evidence they leave behind. ScamTrace helps students investigate suspicious offers by connecting payment destinations, contacts, and narratives. This demonstration uses fictional offers and invented identifiers.
Different names. Shared destination.
Here are six reports. Three differently named offers share one payment destination. Two unrelated offers stay separate, and the duplicate repost adds no independent corroboration. Select the payment node, and ScamTrace shows the original quote from each independent source. Every connection is inspectable. A shared destination is a relationship to investigate, not proof of fraud.
Challenge a connection
Evidence must also be correctable. We will change the payment reading on the Summit report. Watch the graph: three corroborated report pairs become one. The unsupported connections retract immediately. A remaining phone overlap can still appear as a supporting lead. The original source stays preserved, and the edited identifier is clearly attributed to a human correction.
A screenshot becomes reviewable evidence
Now we add a scholarship screenshot. Tesseract reads the image locally in the browser and extracts its transcript. The payment account, phone, and email all need individual review. Adding the screenshot alone creates no new corroborated connections. After checking the payment destination against the original image, we confirm that reading. Only then does it connect to the matching reports. Opening the correction control also reveals the image region supporting the extracted value.
Local AI surfaces narrative leads
ScamTrace also runs a quantized MiniLM model in a browser worker. It compares the stories with identifiers removed, helping surface similar narratives even when the wording changes. These cosine similarity scores are retrieval leads, not fraud probabilities. Neural similarity stays separate from the corroborated evidence graph. Evidence text and screenshots are not uploaded to a server.
Measure the connection policy
The evaluation panel compares our evidence policy with a naive overlap baseline on thirty two labelled fictional reports, or four hundred ninety six pairs. Both find all forty positive pairs. Raw overlap creates two hundred eighty eight false connections; grounded corroboration creates zero on this synthetic dataset. This measures the configured identifier policy. It does not establish real world fraud accuracy or evaluate the neural model.
Keep the chain of evidence
Finally, export the case. Original transcripts, screenshots, identifier readings, corrections, and supporting connections travel together in a portable file. Personal evidence is kept separate from the fictional demo and stored in this browser. ScamTrace is live on Vercel. It helps people follow a defensible chain of evidence, challenge mistakes, and decide what deserves further investigation.