SeniorCare

Inspiration

SeniorCare was inspired by how difficult many government websites are for seniors and other non-technical users. Services like license renewal, Medicare enrollment, and Social Security support often require navigating outdated interfaces, dense instructions, and multi-step portals that are hard to understand. We wanted to build a calmer, more supportive interface that helps people complete real tasks without having to struggle through every part of the system alone.

What it does

SeniorCare is a senior-friendly government assistance platform that combines retrieval and automation. It can:

• answer government-service questions using indexed source material
• retrieve relevant guidance for DMV, Medicare, and similar tasks
• use Amazon Nova Act to assist with live browser navigation
• autofill approved non-sensitive fields in supported workflows
• stop and return control to the user when login, CAPTCHA, payment, or final confirmation is required The goal is not full automation. The goal is safe, practical assistance.

How we built it

We built the backend with FastAPI and organized it into two main flows:

• RAG question answering
• live browser automation

For the RAG side, we:

• gathered government guidance content into local text documents
• indexed that content into OpenSearch
• created a retrieval service that searches for relevant context
• passed the retrieved context into Nova Lite via Amazon Bedrock to generate simpler answers

For the automation side, we:

• created a task-based structure for supported website actions
• defined a BMV renewal workflow in JSON
• built a Nova Act runner that loads a task definition, opens the site, and follows safety rules
• added structured states like COMPLETED, FAILED, and HUMAN_ASSIST_REQUIRED

We also kept the system modular so retrieval, Bedrock inference, and automation logic were separated instead of mixed into one file.

Challenges we ran into

A major challenge was environment setup and model integration. We ran into issues with:

• Bedrock authentication and credential formats
• Python environment mismatches
• Playwright and Nova Act browser startup
• SSL and browser runtime issues
• fragile real-world website navigation

We also learned that government websites are much harder to automate reliably than controlled demo sites. Even when the browser session works, the actual renewal path may be unclear, hidden behind multiple layers, or require early human intervention. Another challenge was voice integration. We originally planned around Nova Sonic, but time and quota limits forced us to rethink how much of the voice layer we could realistically complete for the demo.

Accomplishments that we're proud of

We are proud that we built a working foundation that connects retrieval and live automation in one project.

Some key accomplishments:

• a FastAPI backend with both /query and /automate flows
• OpenSearch-based retrieval connected to Nova Lite responses
• a real Nova Act automation pipeline with safety boundaries
• structured task configuration for browser workflows

We are also proud that we stayed realistic about safety. Instead of claiming full autonomous completion of government transactions, we designed the system to pause when a human is needed.

What we learned

We learned that useful AI systems are not just about choosing a strong model. They depend on:

• clean architecture
• clear task boundaries
• grounded retrieval
• careful handling of interruptions
• reliable runtime setup

We also learned that human handoff is an important product feature, not a weakness. In this kind of workflow, trust matters more than pretending the AI can do everything alone.

What's next

Next, we want to:

• support more workflows beyond BMV renewal
• add real user profile storage for autofill
• expand the RAG dataset with more government sources
• improve frontend usability for seniors
• add a more complete voice interface
• support session continuation after human intervention

Longer term, we want SeniorCare to become a guided gateway for real public-service tasks, where users can understand what they need, get to the right page faster, and complete the safe parts with far less frustration.

Built With

• amazon-bedrock
• css
• fastapi
• nova
• opensearch
• python
• react