Inspiration

Household deadlines arrive scattered across letters, PDFs and emails. An insurance renewal, a school fee and a visa application can each seem manageable—until they land close together.

We wanted a household assistant that could answer three practical questions: What needs my attention? Where did that date come from? Can you remind me?

Duebook explores that experience through a text-based Alexa+ simulation, with a working AWS backend.

What it does

Duebook turns household notices into saved deadlines with supporting evidence.

When a document says “renew within 30 days of receipt” but gives no receipt date, Duebook asks for the missing information before saving. The saved record separates the document quote, the user’s clarification and the date calculation.

It also flags past dates for confirmation, handles duplicate uploads, and checks for nearby deadlines that may compete for attention.

Reminders are optional and require approval. Users can preview the message and schedule, enable reminders for one deadline, snooze the next reminder, or cancel the plan. Approved reminders run on AWS even when the local computer is asleep.

How we built it

Amazon Bedrock Nova Lite extracts structured information from documents. Application code validates the output and calculates supported relative dates. A Strands agent uses MCP read tools to answer questions about saved deadlines and check for conflicts.

Deadline records live in a local Markdown vault with YAML metadata, keeping the evidence readable and persistent across conversations.

For reminders, EventBridge Scheduler invokes a Lambda worker. A private S3 object holds approved reminder snapshots and delivery state, and Amazon SES sends the emails. Uncertain send outcomes pause delivery rather than risk an automatic duplicate.

The interface keeps upload feedback beside the document controls and shows evidence and reminder actions on the deadline card.

Challenges we faced

The hardest problem was distinguishing a plausible answer from a supported one.

An early extraction invented an anchor date for a relative deadline. We added deterministic receipt-date handling and preserved the user’s clarification separately from the document evidence.

Conversational paraphrasing could also blur the distinction between a supplied fact and a fact found in the notice. Direct provenance answers now use the saved record fields.

User testing revealed practical interface problems: upload results appeared too far from the upload button, past dates were saved too readily, and duplicate uploads produced technical errors. We moved feedback beside the controls, added overdue confirmation, and linked duplicates to existing records.

During final rehearsal, malformed model output exposed another failure case. We added one bounded regeneration attempt while retaining validation: invalid results cannot write records.

What we learned

A useful household assistant needs more than fluent answers. It needs persistent records, visible evidence, clear uncertainty and explicit approval before taking action.

Confidence does not independently verify a user-supplied fact. A successful email API call does not prove inbox delivery. A cloud reminder must remain understandable and cancellable after the conversation ends.

Those distinctions shaped both the architecture and the interface.

Accomplishments that we're proud of

Duebook connects document understanding, clarification, persistent evidence, conversational planning and approved cloud reminders in one working flow.

The offline test suite passed 145 tests, with one optional live test skipped. Live document and browser rehearsals exercised extraction, clarification and saving. Scheduled email delivery was also tested with the local server stopped, and receipt was confirmed by the user.

What’s next for Duebook

Next steps include OCR for scanned documents, calendar integration, stronger reminder delivery monitoring and multi-household access controls. We would also explore direct Alexa+ integration when suitable developer access is available.

Prototype scope

This is a local, text-based Alexa+ simulation, not an integration with Alexa’s backend. The demo uses synthetic household documents and creator narration over edited captures of the working application.

Scanned PDFs require OCR, which is not yet implemented. Email delivery currently uses a verified test recipient. Approved cloud reminders store snapshots, so changes and cancellations must be made online through the app.

Source code, setup instructions and a detailed developer friction log are available in the repository.

Built With

• agents
• amazon
• amazon-web-services
• bedrock
• boto3
• context
• eventbridge
• html
• javascript
• lambda
• model
• nova
• protocol
• python
• s3
• scheduler
• ses
• strands