Inspiration

The main inspiration for our project is that we all frequently get scam calls, and seeing Photon's voice technology, we decided on trying to solve this problem. At first, it started as a basic scam call protection agent. But once we realized the potential of our pipeline, we decided to broaden it to handle regular, everyday phone calls, from all people.

What it does

Sift is an AI agent that responds to all callers, and has a short conversation with them to determine if they are a trustable caller. It factors in phone number, whether they're a contact, whether they know your name, if they were flagged previously, and more. The result is an AI agent that you setup once, and never touch again. It invisibly filters out all scam calls, routes trusted people to you, and sends a text message to you instead when you're busy.

How we built it

We used Next.js and Tailwind for the frontend. For the backend, we used Supabase, ElevenLabs, Anthropic API, and Photon. Hosted on Vercel.

Challenges we ran into

It is easy for the work of different group members to overlap, and for excessive files to be created. We had to trim down a lot of unused files from our git repo after merging and reorganize our codebase for legibility. We also had to try various system prompts to enable our call handler to behave as we expect.

Accomplishments that we're proud of

One of the features we're most proud of is our seamlessness. After installation, using Sift is as easy as giving out your Sift number instead of your personal number.

What we learned

One of the biggest things we learned was to prep and plan more before a hackathon, especially since this was pretty much our first experience with prompting and shipping something this complex this quickly. With our original plan, we thought we could work on things completely separately, and integrate at the end. But very early on, we realized that without a clear framework, code overlap and different structures cause merge conflicts very easily.