SIM Continuity Guard - AI-powered pre-issuance risk check

The problem: When a telecom operator reassigns a phone number, the digital world doesn't always deactivate in sync - banking apps, government ID platforms, messaging apps, and other services often stay linked to the previous owner long after the line changes hands. The new subscriber can end up receiving someone else's one-time passcodes and notifications, while the previous owner is left exposed to account-takeover risk. This is a documented, quantified problem industry-wide (a Princeton University study found 66% of recycled numbers retain active third-party links), and it's sharper in markets where a phone number doubles as an identity anchor for banking and government services.

What I did: I picked stc, Saudi Arabia's largest telecom operator, as a grounding case and worked it end-to-end the way an AI PM would scope a real initiative:

Sized the opportunity using stc's public subscriber figures (30.3M) combined with industry benchmark rates, explicitly labeling every assumption since real reassignment data isn't public

Mapped the affected users - new subscriber, previous owner, retail agent, and the trust & safety team - and what each needs from a fix

Designed an AI solution: a real-time "SIM Continuity Risk Score" that combines dormancy window, mobile number portability status, identity-match confidence, and request velocity into an explainable 0–100 score, routing each request to auto-approval, step-up verification, or manual review

Defined the data and modeling approach (rules-based MVP → gradient-boosted classifier), with explainability treated as a hard requirement rather than a nice-to-have, since a false positive here delays a real customer's access to connectivity

Named the trade-offs an AI PM actually owns on this kind of problem: fraud reduction vs. added friction, model accuracy vs. fairness across customer segments, and how to design feedback loops so the system improves over time

What I built: Beyond the written case study, I built a working interactive prototype of the risk-scoring logic - adjust the input signals and watch the score, risk band, and recommended action update live with explainable reason codes, the same way the model's output would need to justify itself to a retail agent and a customer.

This is a self-directed portfolio case study, not a commissioned stc engagement - built to demonstrate how I'd approach a real AI product problem from problem framing through to a tangible prototype.