SentinelEdge
Privacy-first, on-device scam detection for live phone calls — built in a single day, now the basis of a paper submitted to EMNLP 2026.
Overview
Built in a single day at the Claude Builder Club Hackathon at ASU, SentinelEdge is a privacy-first, on-device scam detection system for live phone calls. It won 1st place in the AI Tooling track, and the underlying approach is now the basis of a paper submitted to EMNLP 2026.
Problem
The people most vulnerable to phone scams are often the least equipped with technical defenses — and most existing scam-detection approaches require sending call data to the cloud, a privacy tradeoff that's a non-starter for many of the users who need protection most.
Business Context
Built for a hackathon track judged on real-world AI tooling, differentiated specifically on the privacy-first, on-device constraint rather than raw model accuracy — a deliberate positioning bet that paid off in the judging.
Architecture
Live call audio is analyzed entirely on-device by a compressed, fine-tuned small language model, with periodic federated learning updates improving the model over time — no raw call data ever leaves the device.
Tech Stack
Tradeoffs
The core tradeoff: on-device, privacy-preserving inference over a larger cloud-hosted model — accepting some ceiling on raw accuracy and model size in exchange for zero data leaving the device. Given the target users, that's a deliberate and well-argued tradeoff, not a limitation to apologize for.
Lessons Learned
Privacy constraints aren't just a compliance checkbox — here, they were the actual design driver that made the product usable by the people who needed it most. Designing for the constraint first, rather than retrofitting privacy onto a cloud-first design, shaped every other decision in the build.
Research / EMNLP
The on-device detection approach from SentinelEdge has been developed into a paper submitted to EMNLP 2026. Full abstract and link will be added once the paper is public.
Gallery
