Cosmas

Agent Cosmas closes the case on your eligibility.

Comment

Inspiration

We were trying to help somebody we know figure out what benefits they qualify for, and found that they qualified for housing benefits and medicare savings but would have never known unless we had done the analysis. We wanted to build something that actually helps these people without invading their privacy.

What it does

Cosmas is a benefits eligibility AI agent with a detective theme. The user answers nine questions about their day-to-day lifestyle and financials, and it tells you which programs you actually qualify for and how you can get them. There is no account, no tracking, no selling your data, just answers.

How we built it

Our frontend is made in React in TypeScript with Tailwind styling. Cosmas has a detective style aesthetic with typewriter effects, scanlines, gold and black colors, because we wanted to make it feel like solving a mystery so people would actually engage with it. The backend is a decision tree (written in TS) that knows 25+ programs, and the Gemini API (python) runs the chatbot, using FastAPI.. This was deployed on Vercel.

Challenges we ran into

One challenge was nailing down what requirements benefits actually have, because they vary on household size, immigration status, state, etc. We had to dig into HHS poverty guidelines and program documentation to get the thresholds right because oversimplifying would give people false hope, which is worse than not telling them anything. We also ran into another challenge with the UX: how do you ask someone about their income without making them nervous. Also, how do you balance the detective theme and its details with its readability on a phone?

Accomplishments that we're proud of

We built something that genuinely respects people's privacy. Nothing gets stored, the data disappears when you close the tab. We covered 25+ programs across healthcare, housing, and financial help. Also, we were able to improve quality by specifying the market (only California users), refining the recommendations. Also, the conditional logic means we only ask you questions that matter (e.g. we don't ask pregnancy questions if you have already said you are over 65).

What we learned

People care a lot more about privacy than we expected. The moment you tell somebody their data stays local, you win half the battle. We also learned that federal benefits are massively understandable once you break them down. Lastly, we learned that designs matter: the case investigation feel makes it feel human and understanding so the user can naturally trust the program.

What's next for Cosmas

Extending the programs from just California to all 50 US states. | Adding direct application links with pre-filled forms. | Multiple different language (e.g. Spanish, Vietnamese translations) | Adding realtime web scraping to find updates in programs

Share this project:

Updates

Submission history