Veridex AI

Paste a suspicious message or link. Veridex AI returns a risk level, a score, the claims found, and the tactics used against you
Explainable score: every point has a named reason, plus an ML second opinion and the likely manipulation sequence.
Each claim is extracted and marked UNVERIFIED because the message gives no independent evidence.
Verification steps always use an independent channel, never the contact details inside the message.
Legitimacy checklist ("if it can't be confirmed, don't act") and a Do/Don't plan with India's 1930 helpline.
ML classifier on 1,034 held-out messages: 119 scams caught, 12 missed, 4 false alarms (threshold 0.5).
The Veridex pipeline: Claim → Evidence → Verification → Safe decision.

Inspiration

Scams are getting more convincing: fake bank calls, fake job offers, fake delivery fees, "family in trouble" messages. Most scam detectors only say "scam" or "safe". But a person under pressure doesn't need a label. They need to know why something is suspicious, which claims can't be trusted, what to check, and what to do next. We built Veridex AI around one idea: Claim → Evidence → Verification → Safe decision.

What it does

You paste a suspicious message or link, and Veridex AI:

Extracts the claims the message makes and marks each one UNVERIFIED when no independent evidence is given
Detects manipulation tactics: authority impersonation, fear/threat, urgency, secrecy, OTP/credential requests, financial pressure, reward/lure
Shows the attack pattern, for example Authority → Fear → Urgency → Financial action
Gives an explainable risk score, where every point traces to a named factor (no black-box percentage)
Checks links for red flags such as no HTTPS, raw IPs, shortened links, throwaway domains and look-alike characters
Gives verification steps through an independent channel (the official app, the number on your card), never the contact details in the message
Asks "what would need to be true for this to be legitimate?" as a checklist: if it can't be confirmed, do not act
Gives a Do / Don't plan, including India's reporting route (cybercrime.gov.in, helpline 1930)

It never claims certainty. It says things like "multiple high-risk indicators and could not be independently verified."

How we built it

Rule-based engine (Python): weighted tactic detectors, claim extraction, link analysis and the explainable score
ML classifier (scikit-learn): TF-IDF + logistic regression trained on the public SMS Spam Collection (UCI, via Kaggle), shown as a second opinion next to the rule-based score
Streamlit for the interface, Plotly for the risk gauge, deployed on Streamlit Community Cloud

Results (held-out test set, 1,034 messages)

Threshold	Accuracy	Precision	Recall	Missed scams
0.5	98.5%	96.7%	90.8%	12
0.3	97.5%	87.2%	93.9%	8

For security, a missed scam is worse than a false alarm, so we report both operating points.

Challenges

Data mismatch: the public dataset is general SMS spam, not Indian bank, UPI or job scams, so the classifier can score real scams low. We combined it with rule-based checks and label the ML output as a second opinion, not a verdict.
Explainability: we wanted a score people can understand, so every point of risk is tied to a named factor.
Deployment: getting the app running on Streamlit Cloud needed fixes to the Python version and dependencies.

What we learned

Explainability matters as much as accuracy for people who aren't security experts, and honest reporting of limits (and of false negatives) builds more trust than a high number.

What works and what doesn't

Works: rule-based analysis of pasted messages and links, explainable scoring, verification steps, checklist, action plan, and the ML classifier with real test-set metrics. Limits: general SMS dataset, English only, keyword-based rules that can miss new wording, no LLM layer yet, no screenshot or voice input. It is a decision-support tool and cannot confirm whether something is a scam.

What's next

India-specific scam data (UPI, KYC, fake jobs, customs), LLM-based claim extraction, screenshot input, Hindi and Marathi support, a family mode for elderly users, and an "I already acted" recovery panel.

Built With

Submitted to

ForgeHacks Online 2026
- Winner All Participants
- Winner All Teams

Created by

I designed the concept and built the core of Veridex AI. I wrote the rule-based engine that detects manipulation tactics, extracts claims, analyses links and produces the explainable risk score. I trained the scikit-learn classifier on the SMS Spam Collection and evaluated it on a held-out test set at two thresholds. I built the Streamlit interface, with the risk gauge and the verification checklist, and deployed it on Streamlit Community Cloud, which meant fixing Python version and dependency issues. I also wrote the README and put together the demo video.

Shruti Jadhav
tested and refined Veridex AI. I wrote realistic sample scam messages (bank, job and delivery-fee scams) to check the tactic detection and reported cases the rules missed. I helped write the verification steps and the Do/Don’t action plan, including India’s reporting route (cybercrime.gov.in and helpline 1930). I also helped take the demo screenshots and review the project write-up. I learned how an explainable risk score works and why reporting false negatives honestly matters.

Aayush Mahadik