Inspiration

The current state of online shopping assistants lies within coupon trackers, price trackers, and review aggregators. None of them A. look at video reviews or B. actually understand and contextualize the information around them.

What it does

P.R.V.S.A (pronounced P-riv-s-a) (or The Price Review Video Sentiment Aggregator) does what the name says! It look at products on Amazon and uses context from Youtube videos, Google reviews, and price comparisons from other websites to make choices of purchasing products easier for a consumer (like you and me). P.R.V.S.A avoids self sponsorship and webpage bias by looking beyond just from the platform you are using and places where we can gather better opinions.

How we built it

We used Python and FastAPI for our backend resources and threading where possible to make response time for clients as fast as possible. We integrated our client facing side using Google Chrome Extensions. We used local models running on NVIDIA GPUs using cloud compute in Vultr for classification and sentiment analysis of context (like youtube videos), Google Gemini for contextualization of results, smaller Gemini models for quick summarization, and finally Gemini Nano’s API in Google Chrome for faster results for clients. We used deterministic systems and APIs (such as Youtube's Data V3 API, Youtube Transcript Parsing, and SerpAPI) to gather information to give structured context to our models instead of just passing the model a link.

Challenges we ran into

Some of us have not used these APIs and Frameworks before so picking up on the new tools was new to us, but exciting!

Accomplishments that we're proud of

Getting everything integrated, along with actually gathering youtube context quickly. Running local models that perform embeddings of raw data (that was sick).

What we learned

Deeper methods of programmatic interfaces like async functions and streaming. We learned how to used cloud compute and still run services locally to use using SSH tunnels.