Bouncer - AI-Powered Consumer Risk Assessment

Inspiration

In the past decade, the issue of data privacy has become increasingly relevant as more malicious uses of public information have come to light.
We hope to show how public information can instead be used for social good by providing business owners with access control via AI-powered consumer risk assessment.

Users input information like their name, email, city, zip code, and/or facial image, which is put into a public information search via Google Custom Search and summarized through Gemini Flash 2.0 for quick yet detailed inference.
Summaries are then run through Claude Sonnet 4 for safe, thoroughly explained risk-level output with reduced bias.
Database owners can set preferences of what information they would like Bouncer to prioritize for evaluation.
Administrators can send notifications according to their configured risk level cut-offs

Deployment: Vercel using the Expo framework
Front End: React Native with Typescript
Back End:
- Claude:
  - Safe risk level analysis on deep search information + user-customized prompt
- Gemini:
  - Queried through Python Flask
  - Thorough summaries on public information deep search via Google Custom Search API
- Resend:
  - Verified domain emails (bouncer-app.com) sent to database owner via Supabase edge functions
Database:
- Supabase:
  - Edge functions
  - Authentication
  - Public schemas

We encountered query limiting issues with Google's Custom Search JSON API that we worried would harm one of our demos.
We ran into obstacles when using Supabase. Finding workarounds to things like row-level security, finding out how triggers worked, as well as understanding edge function invocation took a bit of troubleshooting.

Impact: Most importantly, we are proud of the potential positivity this product may create. With the increasing need for application security, as well as the widespread malicious use of public information, our idea was founded upon spreading a message–that while public information can be dangerous, it can also be used for good.
Thoroughness: This is by far the most thought-out and fully deployed app we have ever made–with comprehensive activity logs, risk summary emails, and the aesthetic dashboard.
Front end: Creating a visually pleasing dashboard display that connected to all of our other functions (and that we all agreed looked good) was a great Eureka! moment.

The harms and benefits of vast public information: We found out how easy it is to find someone's public information through simple web queries. Though we utilized this data for social good, we all became more cognizant about the extent of public information.
LLM strengths and weaknesses: Though many can serve the same general purposes, finding the specific LLMs that best fit our use cases allowed us to find the contextual bounds of each.
Supabase Functionality: We learned a ton about how Supabase and other databases have increasing meta-editing functionality through edge functions as well as triggers.

npm package deployment: We plan to publish Bouncer as an npm package so that it can be easily added as a form of a log-in page to any application (web, iOS, and Android) in a single line.
Verification of inference results: Though our results were accurate, automatically verifying these results would ensure the credibility of our model's risk level output. Using separate models, increasing the depth of information search, and more could be ways to accomplish this.