Consensi

Create Account Part 1
Create Account Part 2
Profile Settings
Create Account Part 3
Bill Card
Feed

Inspiration

Do you know your Senators? Representatives in the House? Chances are, you probably don't. In our current Political Dynamic, Congress has an abysmal approval rate of around 30%. This distrust often stems from a lack of transparency about what goes on every day in Congress. Americans want to have a window into our democracy. That's why, today, we are presenting Consensi: The website that will help you become a better citizen.

What it does

Consensi presents all current legislation in Congress and utilizes AI-powered tools so that complicated Congressional jargon can be simplified for everyone. We do this through Consensi's neat and intuitive For You Page, highlighting the user's most interested topics in rows and the associated bills with each of them. Users can then further click on each bill to get a simplified summary of it, chat with an LLM about further questions about it, show their support or lack of support for it, and also have the option to connect with their representative to voice their opinion. On the other side, representatives can use Consensi to view the demographic info of voters who support the bill, allowing them to stay informed of their constituents' desires.

How we built it

We used the Congressional API to query all the information about the bills of the current Congress and reformatted the text into a readable format. We then ran this information through the Hugging Face model: facebook/bart-large-mnli, which is a BART Model that is trained using zero-shot classification. This model gave us the advantage in that: You can introduce new labels at test time by just phrasing them as hypotheses - no retraining needed. All this information was saved to a central Supabase database that stores all user, bill, and demographic information. We used a Typescript backend to quickly query and send all the data to the user in the front end. The front end was built using React and TailwindCSS, providing a responsive and modern interface for users to explore legislative data interactively. Each bill entry dynamically displays key metadata such as sponsors, subjects, and recent actions, along with the model-generated topic classifications. We also integrated filtering and search functionalities, allowing users to browse bills by category (e.g., Healthcare, Economy, Civil Rights) or by individual members of Congress.

Challenges we ran into

A challenge we ran into was getting the full text of the bills. The Congressional API does not provide that information on its own, so we had to create the link to the bill from the information we had and get the text from that. We also had to clean and reformat the text so it didn’t have tags in it. Ensuring that our text would be queried fully clean took quite a bit of time, as we did not want any noise to throw off our classification. Another challenge we ran into was implementing the chatbot that allowed you to ask questions about the bill. It was quite a fun challenge to implement my own Chatbot interface and ensure the RAG capabilities worked for the appropriate Bill.

Accomplishments that we're proud of

We’re proud of how many robust features we were able to implement in such a short amount of time. We built our own classification system to efficiently categorize otherwise disorganized bills, making it easy to see what each bill affects, while not relying on any outside model or API. We are proud of our system of endorsements and opposition, as this makes communication and transparency go both ways. Not only can average citizens see what the government is planning, but those in power can see what their constituents support, allowing them to better respond to their needs. Finally, we are proud of our sleek and easy-to-use UI. It presents information in an appealing and informative way. It doesn’t oversimplify the bills, but rather allows people to truly understand them without having to struggle through dense legal language.

What we learned

From this project, we learned a ton about databases and how to organize them so that all the information we need is easily accessible. We had used other databases before, but not Supabase specifically, so we learned how to work within its ecosystem and read up on its documentation. We also used OpenRouter for the first time. It allowed us to really easily change what model we were using, which was great when we decided that Gemini would be a better model for our purposes than Claude. Learning effective version control also proved very challenging but rewarding. We all had experience with git before, but we had to deal with a lot of merge conflicts, especially when multiple members were working on the same files or pushing large feature updates simultaneously. Through this, we learned the importance of clear branching strategies, frequent commits, and detailed pull request reviews to maintain a clean and functional codebase.

What's next for Consensi

We hope that this is just the beginning for Consensi. We are going to try and connect this tool with academics at UMass who can help us shape this tool into something lasting. If we can continue this project in the future, one idea I would be interested in implementing is LLM-powered bill searching through the usage of MCP tools that utilize the embedding search.

Built With

css
docker
milvus
python
react
recharts
sql
superbase
typescript

Submitted to

HackUMass XIII

Created by

I initialized the project. I set up the entire file structure and boilerplate of our site using Cursor. I then implemented the Bill Translation logic. I utilized the Gemini API to translate all bills into simple terms that a 10th grader could read. After setting up bill translations and having them pushed to our database. I began to work on the core backend functionality of our website. I created a series of routes to connect our database and user authentication to the client. I also helped with the deployment of our site on docker.

Alan Achilles
I designed and developed the frontend of the website, including the Login, Feed, "My Bills," User Profile, and Landing pages. Additionally, I implemented user, registration, login, and updates in the backend. I also linked other backend features, including the chatbot model and legislator contact information, to the frontend. During this process, I expanded my knowledge of asynchronous programming, design, NextAuth, and TypeScript.

Nathan Gelfand
I first created the api caller that called the Congressional API and put all of the bills into the database. I created a separate scraper that acquired the text of the bills in plaintext form to be easily fed into our ML categorizer.

I created and managed the database, as well as worked on the backend. This was my first time working Supabase, so I learned all about using the platform and integrating it into the project. I created the database schemas and tables for users, bills, and endorsements. I created the backend logic that saved users and their demographic information, and updated the database when that information was changed. I created the endorsement and opposition logic and connected that to the frontend, altering the database based on user actions. Finally, I used the recharts library to visualize the demographics that were supporting a bill, and displayed that when the bill was selected.

Adam Pochobut
I worked on the AI + ML features of our project! In addition to chat-gpt wrapper esque things like utilizing Gemini Flash 2.5 for creating cost effective but accurate summaries and also in our RAG assisted Chatbots for providing additional user support on understanding the bill's technicalities. I also implemented some more traditional ML features.
For one, I used the hugging face model: facebook/bart-large-mnli to label the categories of each of our bills. This model, which is a BART Model, is trained using zero-shot-classification. This gave us the advantage in that: You can introduce new labels at test time by just phrasing them as hypotheses - no retraining needed. For example in a Traditional model: It learns to map inputs -> fixed label set (e.g., “sports”, “politics”, “health”). If you add a new label (“environment”), it must be retrained with labeled examples.
In this model it can take the sentence: “The patient needs immediate care.” And create the Hypothesis: “The text is about medical assistance.” because the Model still “understands” this new concept even though it wasn’t trained on it. Basically, this model makes decisions using natural language hypotheses similar to how humans interpret meaning. This was beneficial to our project because new topics are introduced into legislature all the time and having a model that is able to just understand Natural Language without needing to be retrained is quite advantageous. Our second implementation was an Embedding, Nearest Neighbor type algorithm search. I implemented this because the rigid categories found from the transformer were very broad and users who wanted to query specific unique issues could have a way to this. To implement this I first created embeddings for each of the texts of the bills storing them in a Milvus database, and then would take an embedding of the user query and return the top 10 nearest neighbors to it utilizing Cosine Similarity. For example, a search for "Bills about Health AI" would create an embedding for that query and then find the bills that are most closely related to it and would share a similar semantic meaning.

If we are able to continue this project in the future I would be interested in implementing LLM powered bill searching through the usage of MCP tools that utilize the embedding search.

Adam Lubomirski
Nathan Gelfand