abc - FraudNot | Devpost

System Architecture

Inspiration

The recent deepfake Zoom scam of high-profile government impersonations inspired us to dig deeper into tackling growing scams.

We decided to build FraudNot in existence with ScamShield to increase our combat effectiveness and help provide additional reassurance against scams.

What it does

The user sent a suspicious message, image, or voice note. FraudNot runs them through a forensic analysis pipeline grounded in real Singapore scam cases and returns a verdict with confidence score, scam type, key indicators, and a recommended action all within Telegram.

It is capable of drafting a police report in the format that SPF expects. This improves the user experience of lodging a police report.

How we built it

It is built on Python-telegram-bot and included Reka Flash via HTTPX for grounded analysis of suspicious scam indicators. We used ElevenLabs Scribe for audio transcription which is passed to Reka Flash for analysis. We added RAG engine as existing database to guide Reka analysis to improve accuracy and included a docker for ease of production.

Challenges we ran into

The Reka API key did not give us access to the Reka Core model. We had to think of another solution to break the ambiguity score.

Another challenge included Reka's speech-to-text server not working as intended. We had to rely on another AI model for the transcription. But this also reduces our bot dependency on 1 specific AI model.

Accomplishments that we're proud of

The bot is able to run as intended with low false positives and false negatives and relatively high accuracy.

We were able to utilize what we learned from both the Git workshop and the Reka demo to streamline our project.

What we learned

We learned how to use different tech stacks such as using Reka Flash and creating a docker to develop and improve on our current model. We also learnt how to implement github for collaborative effort.