Libra

Inspiration

We personally know many immigrants who do not speak English natively and are low-income, but they are either unaware of such aid programs or feel anxious applying to them given the complex English terminologies used. We wanted to build a platform and a solution to ease them into applying for the benefits they very much deserve.

What it does

Libra is split into 2 parts, the web app and the extension. The web app first provides a simple introduction to our service, where users tell us their basic information and profile, none of which are stored as data anywhere. Based on the information of their profile, the model then finds suitable aid programs catered to your needs and grounds it based on official sources.

How we built it

We built Libra using a FastAPI backend to handle the logic and a React frontend for the user interface. For the intelligence layer, we utilized OpenDeepSearch integrated with LiteLLM to allow our system to cross-reference real-time government databases and official PDF manuals. We used OpenRouter (Google Gemini 2.0 Flash) to power our multilingual reasoning engine, ensuring that the eligibility checks were grounded in actual policy. The browser extension was developed using the Chrome Extension API, designed to "read" complex application forms and provide real-time, plain-language guidance to the user as they type.

Challenges we ran into

There were significant differences in the quality of responses across different languages; for instance, Somali and Vietnamese responses were initially much less specific than English. We solved this through iterative prompt engineering, forcing the model to "chain-of-thought" reason in English first before translating the final, high-quality logic into the target language. We also struggled with Windows environment compatibility during development, specifically with C++ build dependencies for our search tools, which required us to pivot our environment strategy mid-hackathon.

Accomplishments that we're proud of

We managed to get a working demo on a real website that many people use, and it will definitely be beneficial to many people. We managed to include a lot of different communities of people. For instance, we used a font that is readable specially for low-vision people, ensuring that they are not excluded from being able to apply to these aid programs. Furthermore, we are proud of text-to-speech and speech-to-text support for 8 different languages, ensuring that no immigrants are excluded.

What we learned

We learned that the biggest barrier to government aid isn't just a lack of information, but the administrative burden of jargon and hostile form design. Technically, we learned how to ground LLM responses in specific, external web sources to prevent hallucinations, which is a critical requirement when dealing with legal and financial eligibility. We also gained a deeper understanding of local-first data privacy, realizing that users are much more willing to engage with AI when they know their sensitive information never leaves their browser's local storage.

What's next for Libra

we would implement a formal linguistic audit by having multilingual human evaluators compare English responses against target language outputs. This would ensure that the nuance and specificity of the guidance remain equal across all languages, preventing a "quality gap" for non-English speakers. Furthermore, we would expand Libra’s functionality to include an AI-driven interview preparation module. Many programs, such as CalFresh, require a follow-up interview with a caseworker, and hence our tool would allow users to practice these conversations, reducing the anxiety and potential communication barriers that often lead to benefits being denied at the final stage.

Built With

• deepseek
• fastapi
• groq
• jina
• next.js
• openrouter
• python
• stt
• tts
• typescript
• uvicorn