NavBen

Inspiration

Applying for government assistance shouldn't feel like a maze. For low-income and vulnerable families, navigating 20-page legal forms is a high-stress barrier that frequently blocks access to food, housing, and healthcare. We were inspired to help target users such as vulnerable families, low-income parents, elderly citizens, and immigrants facing sudden financial crises who often struggle with language barriers, text literacy, and the anxiety of clinical questionnaires. A simple formatting error on paperwork can lead to rejection after months of waiting, so we set out to build a system that is compassionate, accurate, and accessible.

What it does

Navben is an intelligent assistant that replaces endless paperwork with a warm, natural conversation. Citizens can interact with the system via text chat or a real-time voice call in over 70 languages. As the user explains their situation (e.g., "I lost my job and have two kids"), the system listens, extracts the relevant profile variables, and runs an instant eligibility check. Instead of handing the user another clinical form to fill out, Navben generates a clean, personalized, step-by-step document checklist (e.g., "Please upload your ID and your last pay stub") that empowers the citizen to confidently secure the support they deserve.

How we built it

We built Navben using a dual-brain Neuro-Symbolic architecture to separate language processing from strict mathematical logic:

• The Neuro Brain: We used Gemini 3.1 Flash-Lite for text chat intake, Natural Language Processing (NLP), and policy data extraction. For the multilingual voice assistant, we implemented Gemini 2.5 Flash via LiveKit Cloud for native real-time audio-to-audio streaming. These models translate the input and extract structured profile variables (e.g., location: "Abuja", age: 25, employment: "unemployed").
• The Symbolic Brain: We built a deterministic, rule-based TypeScript evaluation engine that takes the variables extracted by the LLMs and runs them against strict program logic trees saved in a Firebase database.

Challenges we ran into

The biggest challenge was mitigating the risk of AI hallucinations in eligibility math. Standard LLMs frequently make mathematical reasoning errors. In a high-stakes environment like government assistance, if an LLM acts as the sole decision-maker, a single hallucination could wrongfully deny food stamps to a hungry family or lead to fraud penalties by approving someone who is legally ineligible. Overcoming this required strict discipline to not let the LLM evaluate rules autonomously, but rather strictly confine it to variable extraction.

Additionally, ensuring access equity was a major challenge. We realized a text-based chatbot still presented barriers for text-illiterate or non-English speaking applicants, requiring us to successfully implement real-time, translated voice streaming.

Accomplishments that we're proud of

We are incredibly proud of achieving a 0% hallucination rate for eligibility verdicts. By decoupling conversational NLP from the deterministic TypeScript evaluator, our system is mathematically flawless when evaluating the law.

We are also proud of our strict Human-in-the-Loop design. The AI never submits or decides anything autonomously. It hands control back to the citizen to upload their physical documents, and then routes the completed checklist to a human caseworker. The AI has no authority to distribute funds; a human caseworker always logs into their CRM dashboard to click the final "Approve Benefits" button.

What we learned

We learned that in high-stakes domains, AI should guide and assist, not finalize or deny. Separating conversational data extraction from hardcoded, rule-based evaluation is the safest and most reliable way to deploy AI in government and legal services. Furthermore, we learned that giving users a conversational voice interface rather than a blank form drastically lowers the cognitive load and barrier to entry for vulnerable populations.

What's next for NavBen

Looking ahead, our next step for Navben is to integrate directly with state government APIs to securely auto-verify uploaded documents like utility bills and IDs, reducing the caseworkers' manual review time even further. We also plan to expand access by deploying the voice agent natively over standard phone calls (telephony) and WhatsApp, ensuring that anyone with a basic mobile phone no matter their language or location can easily access the public benefits they deserve.

Built With

• gemini
• javascript
• livekit
• python
• react
• vercel