Inspiration
Phone scams are an everyday danger in Bangladesh. Fake bKash and Nagad agents, bank OTP phishing calls, and fake lottery wins cost real victims real money — and most of them have nowhere to report what happened in a way that actually protects the next person. We wanted to build something that does two things at once: warn people before they lose money, and turn victims' experiences into protection for others, instead of letting each scam be a dead end.
What it does
Scam Call Shield has two connected flows:
- Check a Call — paste in (or speak) what was said on a suspicious call, and NVIDIA Nemotron instantly analyzes it for scam patterns — urgency, impersonation, OTP/PIN phishing, fake prizes — and returns a confidence score, the scam type, and a plain-language warning.
- Report a Scam — after verifying their phone number with a simple OTP (no government ID required, to keep the barrier low while filtering spam), a victim can submit what happened to them. Every verified report joins a community database that future checks are compared against — so the system gets smarter the more people use it.
How we built it
The backend is a FastAPI app that calls NVIDIA Nemotron through
Nebius Token Factory's OpenAI-compatible API. Nemotron models on
Nebius are reasoning models — they can return their answer in a
reasoning_content field instead of (or alongside) content — so the
client explicitly handles both and asks the model to end with a
parseable JSON verdict. A lightweight retrieval step feeds a summary of
prior verified community reports into the prompt as context, so
detections are grounded in real local scam patterns rather than generic
knowledge. The frontend is a single self-contained HTML page (Bangla
UI, since that's the language our real users speak) that talks to the
backend over a simple REST API, with the browser's built-in speech
recognition for hands-free transcript input.
Challenges we ran into
The biggest challenge wasn't technical — it was access. Nebius's billing system didn't initially support Bangladesh in its "country of residence" list, which blocked API key creation entirely. We worked through it by reaching out to Nebius support on Discord and the hackathon organizers, redeeming the hackathon's Token Factory promo credit, and eventually getting a verified card through to complete billing. On the technical side, we also deliberately chose not to claim NVIDIA Riva/Parakeet for speech-to-text, since those speech models aren't currently offered on Nebius Token Factory — we used the browser's Web Speech API instead and kept all AI reasoning on NVIDIA Nemotron, so every claim in this submission is something you can actually verify by running the code.
What we learned
That "working demo" is a much higher bar than "working code on my machine" — getting from a model that returns good answers in testing to a model that reliably returns parseable answers in production (Nemotron is a reasoning model, and its output format needed real handling, not just a hopeful regex) took real iteration.
What's next
Real SMS delivery for OTPs (currently demo-mode only), vector-embedding search over the report database for better retrieval, and a browser extension or call-screening integration so Scam Call Shield can warn people on real incoming calls, not just pasted transcripts.
Log in or sign up for Devpost to join the conversation.